From 97cdbd061df2e48d49355c095581f5a710363960 Mon Sep 17 00:00:00 2001 From: Seongho Bae Date: Tue, 8 Sep 2026 14:00:30 +0900 Subject: [PATCH 01/12] test(scanner): fail closed on plugin s3 writes and containerapp up Hook and manifest aws s3 sync/cp and az containerapp up must fail closed. aws s3 ls, comments, echo lookalikes, and README wording stay inventory. Relates to #1099. --- tests/test_claude_plugin_object_store.py | 198 +++++++++++++++++++++++ 1 file changed, 198 insertions(+) create mode 100644 tests/test_claude_plugin_object_store.py diff --git a/tests/test_claude_plugin_object_store.py b/tests/test_claude_plugin_object_store.py new file mode 100644 index 00000000..3e9dda14 --- /dev/null +++ b/tests/test_claude_plugin_object_store.py @@ -0,0 +1,198 @@ +"""Hook aws s3 writes and az containerapp up fail closed; ls stays inventory.""" + +from __future__ import annotations + +import json +from pathlib import Path + +from appguardrail_core.claude_plugin_detector import ( + _collect_plugin_hits, + build_claude_plugin_scan_receipt, + inspect_claude_plugin_file, + inventory_claude_plugin_capabilities, +) + + +_PINNED_COMMIT = "a727be1c7bd6064419b6f60d71993a19198adc17" +_S3_RULE = "claude-plugin-aws-s3-write-command" +_CONTAINERAPP_RULE = "claude-plugin-az-containerapp-up-command" +_AWS_DEPLOY_RULE = "claude-plugin-aws-deploy-command" +_AZ_DEPLOY_RULE = "claude-plugin-az-deploy-command" +_SECRET = "sk-object-must-not-leak" +_BIDI = "\u202e" +_THIS_CLASS = frozenset({_S3_RULE, _CONTAINERAPP_RULE}) + + +def _write_json(path: Path, payload: dict) -> None: + """Write one JSON document under ``path``.""" + path.parent.mkdir(parents=True, exist_ok=True) + path.write_text(json.dumps(payload, indent=2) + "\n", encoding="utf-8") + + +def _licensed_plugin(root: Path, hook_body: str = "#!/bin/sh\necho hello\n") -> Path: + """Write a pinned licensed plugin with one declared shell hook.""" + _write_json( + root / ".claude-plugin" / "plugin.json", + { + "name": "safe-plugin", + "version": "1.0.0", + "source": { + "source": "github", + "repo": "example/safe-plugin", + "ref": _PINNED_COMMIT, + }, + "hooks": {"PreToolUse": [{"command": "hooks/session.sh"}]}, + }, + ) + hook = root / "hooks" / "session.sh" + hook.parent.mkdir(parents=True, exist_ok=True) + hook.write_text(hook_body, encoding="utf-8") + hook.chmod(0o755) + (root / "LICENSE").write_text("MIT\n", encoding="utf-8") + return root + + +def _hits(root: Path, rule_id: str): + """Return receipt-path hits for one rule identity.""" + return [hit for hit in _collect_plugin_hits(root) if hit.rule_id == rule_id] + + +def test_hook_aws_s3_sync_fails_admission(tmp_path: Path) -> None: + """``aws s3 sync`` on a hook is object-store write authority.""" + root = _licensed_plugin(tmp_path, "#!/bin/sh\naws s3 sync ./dist s3://bucket/app\n") + hits = _hits(root, _S3_RULE) + receipt = build_claude_plugin_scan_receipt(root) + inventory = inventory_claude_plugin_capabilities(root) + + assert hits + assert all(hit.snippet == "aws s3 sync" for hit in hits) + assert receipt.scan_result == "fail" + assert _S3_RULE in receipt.finding_summary + assert _CONTAINERAPP_RULE not in receipt.finding_summary + assert _AWS_DEPLOY_RULE not in receipt.finding_summary + assert inventory["deployment_write"] is True + + +def test_hook_aws_s3_cp_fails_admission() -> None: + """``aws s3 cp`` is the same object-store write class.""" + body = "#!/bin/sh\naws s3 cp artifact.tgz s3://bucket/app.tgz\n" + hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) + assert any(hit.rule_id == _S3_RULE and hit.snippet == "aws s3 cp" for hit in hits) + + +def test_hook_az_containerapp_up_fails_admission(tmp_path: Path) -> None: + """``az containerapp up`` on a hook is Azure write authority.""" + root = _licensed_plugin(tmp_path, "#!/bin/sh\naz containerapp up --name app\n") + hits = _hits(root, _CONTAINERAPP_RULE) + receipt = build_claude_plugin_scan_receipt(root) + + assert hits + assert all(hit.snippet == "az containerapp up" for hit in hits) + assert receipt.scan_result == "fail" + assert _CONTAINERAPP_RULE in receipt.finding_summary + assert _AZ_DEPLOY_RULE not in receipt.finding_summary + + +def test_aws_s3_ls_stays_inventory(tmp_path: Path) -> None: + """``aws s3 ls`` stays inventory, not this class.""" + root = _licensed_plugin(tmp_path, "#!/bin/sh\naws s3 ls\n") + receipt = build_claude_plugin_scan_receipt(root) + + assert _hits(root, _S3_RULE) == [] + assert _THIS_CLASS.isdisjoint(receipt.finding_summary) + assert receipt.scan_result == "pass" + + +def test_comment_and_echo_object_store_are_not_this_class(tmp_path: Path) -> None: + """Unquoted comments and echo lookalikes are not executable s3 writes.""" + root = _licensed_plugin( + tmp_path, + '#!/bin/sh\n# aws s3 sync ./dist s3://bucket\necho "az containerapp up"\n', + ) + receipt = build_claude_plugin_scan_receipt(root) + assert _THIS_CLASS.isdisjoint(receipt.finding_summary) + assert receipt.scan_result == "pass" + + +def test_readme_s3_sync_is_not_this_class(tmp_path: Path) -> None: + """README object-store wording is repository guidance, not a hook command.""" + root = _licensed_plugin(tmp_path) + (root / "README.md").write_text("aws s3 sync ./dist s3://bucket/app\n", encoding="utf-8") + receipt = build_claude_plugin_scan_receipt(root) + inventory = inventory_claude_plugin_capabilities(root) + + assert _hits(root, _S3_RULE) == [] + assert receipt.scan_result == "pass" + assert inventory["deployment_write"] is True + + +def test_echo_then_real_s3_sync_still_fails() -> None: + """``echo done && aws s3 sync`` still runs the object-store write.""" + hits = inspect_claude_plugin_file( + "session.sh", + "hooks/session.sh", + '#!/bin/sh\necho "done" && aws s3 sync ./dist s3://bucket\n', + ) + assert any(hit.rule_id == _S3_RULE and hit.snippet == "aws s3 sync" for hit in hits) + + +def test_snippets_are_command_labels_not_secrets(tmp_path: Path) -> None: + """Snippets name the CLI command and omit secrets and bidi.""" + body = f"#!/bin/sh\naws s3 cp secret.bin s3://bucket/{_SECRET}{_BIDI}\n" + root = _licensed_plugin(tmp_path, body) + hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) + s3_hits = [hit for hit in hits if hit.rule_id == _S3_RULE] + payload = json.dumps(build_claude_plugin_scan_receipt(root).as_dict()) + + assert s3_hits + for hit in s3_hits: + assert hit.snippet == "aws s3 cp" + assert _SECRET not in hit.snippet + assert _BIDI not in hit.snippet + assert _SECRET not in hit.message + assert _SECRET not in payload + assert _BIDI not in payload + + +def test_plugin_manifest_containerapp_up_fails_admission(tmp_path: Path) -> None: + """A plugin.json command string that runs containerapp up is that class.""" + root = _licensed_plugin(tmp_path) + manifest = json.loads( + (root / ".claude-plugin" / "plugin.json").read_text(encoding="utf-8") + ) + manifest["hooks"] = { + "PostToolUse": [{"command": "az containerapp up --name app"}], + } + _write_json(root / ".claude-plugin" / "plugin.json", manifest) + receipt = build_claude_plugin_scan_receipt(root) + assert _hits(root, _CONTAINERAPP_RULE) + assert receipt.scan_result == "fail" + assert _CONTAINERAPP_RULE in receipt.finding_summary + + +def test_manifest_prose_and_reporting_are_not_this_class(tmp_path: Path) -> None: + """Manifest prose and reporting-only values are not object-store writes.""" + root = _licensed_plugin(tmp_path) + manifest_path = root / ".claude-plugin" / "plugin.json" + manifest = json.loads(manifest_path.read_text(encoding="utf-8")) + manifest["description"] = "operators may later run aws s3 sync" + manifest["hooks"] = { + "PostToolUse": [{"command": 'echo "az containerapp up"'}], + } + _write_json(manifest_path, manifest) + receipt = build_claude_plugin_scan_receipt(root) + + assert _THIS_CLASS.isdisjoint(receipt.finding_summary) + assert receipt.scan_result == "pass" + + +def test_cloudformation_deploy_stays_the_aws_deploy_class() -> None: + """CloudFormation deploy without s3/containerapp stays the aws-deploy class.""" + hits = inspect_claude_plugin_file( + "session.sh", + "hooks/session.sh", + "#!/bin/sh\naws cloudformation deploy --stack-name app\n", + ) + rule_ids = {hit.rule_id for hit in hits} + assert _AWS_DEPLOY_RULE in rule_ids + assert _THIS_CLASS.isdisjoint(rule_ids) From 8ca0628d96d5dc4db3a78de6e620a1ab4049b0e9 Mon Sep 17 00:00:00 2001 From: Seongho Bae Date: Tue, 8 Sep 2026 14:04:27 +0900 Subject: [PATCH 02/12] feat(scanner): reject plugin s3 writes and containerapp up Fail closed on executable aws s3 sync/cp as claude-plugin-aws-s3-write-command and az containerapp up as claude-plugin-az-containerapp-up-command. aws s3 ls, comments, and echo lookalikes stay inventory. Relates to #1099. --- .../1099-claude-plugin-supply-chain.md | 3 + appguardrail_core/claude_plugin_detector.py | 84 ++++++++++++++++++- docs/TRACEABILITY.md | 2 +- docs/sast-dast-rule-research.md | 5 +- tests/test_claude_plugin_cloud_deploy.py | 3 + tests/test_claude_plugin_hosted_deploy.py | 3 + tests/test_claude_plugin_object_store.py | 2 + 7 files changed, 97 insertions(+), 5 deletions(-) diff --git a/CHANGELOG.d/1099-claude-plugin-supply-chain.md b/CHANGELOG.d/1099-claude-plugin-supply-chain.md index 44520912..0e5806de 100644 --- a/CHANGELOG.d/1099-claude-plugin-supply-chain.md +++ b/CHANGELOG.d/1099-claude-plugin-supply-chain.md @@ -150,6 +150,9 @@ `claude-plugin-aws-deploy-command`. ``gcloud run|app|functions deploy`` fails as `claude-plugin-gcloud-deploy-command`. ``az webapp deploy`` fails as `claude-plugin-az-deploy-command`. + ``aws s3 sync`` and ``aws s3 cp`` fail as + `claude-plugin-aws-s3-write-command`. ``az containerapp up`` fails as + `claude-plugin-az-containerapp-up-command`. Hook comments and ``echo``/``printf`` lookalikes are not those classes. ``terraform plan``, ``helm list``, diff --git a/appguardrail_core/claude_plugin_detector.py b/appguardrail_core/claude_plugin_detector.py index 0312fa38..bfe4deec 100644 --- a/appguardrail_core/claude_plugin_detector.py +++ b/appguardrail_core/claude_plugin_detector.py @@ -27,7 +27,9 @@ hosted-deploy command findings. Hook or manifest ``aws cloudformation deploy``, ``aws deploy create-deployment``, ``gcloud run|app|functions deploy``, and ``az webapp deploy`` fail closed as cloud-deploy command -findings. Unquoted ``#`` comments and +findings. Hook or manifest ``aws s3 sync``, ``aws s3 cp``, and +``az containerapp up`` fail closed as object-store and Container Apps +writes. Unquoted ``#`` comments and ``echo``/``printf``/``print`` lookalikes are not that class. ``terraform plan``, ``helm list``, ``vercel ls``, ``fly status``, ``aws s3 ls``, ``gcloud config list``, and ``az account show`` @@ -296,6 +298,16 @@ "Azure web app is write authority. Remove the command. " "[CWE-269 - Improper Privilege Management]" ) +CLAUDE_PLUGIN_AWS_S3_WRITE_COMMAND_MESSAGE: Final = ( + "Claude plugin hook or manifest runs aws s3 sync or aws s3 cp. " + "Copying objects into a bucket is write authority. Remove the " + "command. [CWE-269 - Improper Privilege Management]" +) +CLAUDE_PLUGIN_AZ_CONTAINERAPP_UP_COMMAND_MESSAGE: Final = ( + "Claude plugin hook or manifest runs az containerapp up. Publishing " + "a Container Apps revision is write authority. Remove the command. " + "[CWE-250 - Execution with Unnecessary Privileges]" +) CLAUDE_PLUGIN_DOCKER_SOCKET_MESSAGE: Final = ( "Claude plugin hook reaches the host Docker socket. Socket access is host " "control, not an image push. Remove the socket bind and keep builds " @@ -431,6 +443,11 @@ re.IGNORECASE, ) _AZ_DEPLOY_COMMAND = re.compile(r"\baz\s+webapp\s+deploy\b", re.IGNORECASE) +_AWS_S3_WRITE_COMMAND = re.compile(r"\baws\s+s3\s+(?Psync|cp)\b", re.IGNORECASE) +_AZ_CONTAINERAPP_UP_COMMAND = re.compile( + r"\baz\s+containerapp\s+up\b", + re.IGNORECASE, +) _REPORTING_BUILTINS: Final = frozenset({"echo", "printf", "print"}) _FIRST_SHELL_TOKEN = re.compile(r"\s*([A-Za-z0-9_./+-]+)") _DOCKER_SOCKET = re.compile( @@ -668,8 +685,9 @@ re.compile( r"\b(?:kubectl\s+apply|terraform\s+apply|helm\s+install|" r"vercel\s+deploy|fly(?:ctl)?\s+deploy|docker\s+push|" - r"aws\s+(?:cloudformation\s+deploy|deploy\s+create-deployment)|" - r"gcloud\s+(?:run|app|functions)\s+deploy|az\s+webapp\s+deploy)\b", + r"aws\s+(?:cloudformation\s+deploy|deploy\s+create-deployment|s3\s+(?:sync|cp))|" + r"gcloud\s+(?:run|app|functions)\s+deploy|az\s+webapp\s+deploy|" + r"az\s+containerapp\s+up)\b", re.IGNORECASE, ), ), @@ -928,6 +946,8 @@ def inspect_claude_plugin_file( hits.extend(_aws_deploy_command_hits(content, manifest=manifest)) hits.extend(_gcloud_deploy_command_hits(content, manifest=manifest)) hits.extend(_az_deploy_command_hits(content, manifest=manifest)) + hits.extend(_aws_s3_write_command_hits(content, manifest=manifest)) + hits.extend(_az_containerapp_up_command_hits(content, manifest=manifest)) hits.extend(_docker_socket_hits(content)) hits.extend(_browser_profile_hits(content)) hits.extend(_credential_store_hits(content)) @@ -1999,6 +2019,64 @@ def _az_deploy_command_hits( return () +def _aws_s3_write_command_hits( + content: str, *, manifest: bool = False +) -> tuple[PluginHit, ...]: + """Return ``aws s3 sync``/``cp`` findings with a command label, not URIs. + + Args: + content: Hook or manifest text. + manifest: When true, only structural command values are scanned. + + Returns: + One hit for executable ``aws s3 sync`` or ``aws s3 cp``. + ``aws s3 ls``, comments, and echo lookalikes are not this class. + """ + for source, first_line in _hosted_command_sources(content, manifest=manifest): + match = _executable_command_match(source, _AWS_S3_WRITE_COMMAND) + if match is None: + continue + verb = match.group("verb").lower() + return ( + PluginHit( + rule_id="claude-plugin-aws-s3-write-command", + line=first_line + source[: match.start()].count("\n"), + snippet="aws s3 " + verb, + message=CLAUDE_PLUGIN_AWS_S3_WRITE_COMMAND_MESSAGE, + ), + ) + return () + + +def _az_containerapp_up_command_hits( + content: str, *, manifest: bool = False +) -> tuple[PluginHit, ...]: + """Return ``az containerapp up`` findings with a command label, not names. + + Args: + content: Hook or manifest text. + manifest: When true, only structural command values are scanned. + + Returns: + One hit when executable ``az containerapp up`` is present. + ``az account show``, comments, and echo lookalikes are not this + class. + """ + for source, first_line in _hosted_command_sources(content, manifest=manifest): + match = _executable_command_match(source, _AZ_CONTAINERAPP_UP_COMMAND) + if match is None: + continue + return ( + PluginHit( + rule_id="claude-plugin-az-containerapp-up-command", + line=first_line + source[: match.start()].count("\n"), + snippet="az containerapp up", + message=CLAUDE_PLUGIN_AZ_CONTAINERAPP_UP_COMMAND_MESSAGE, + ), + ) + return () + + def _dynamic_eval_hits(content: str) -> tuple[PluginHit, ...]: """Return findings for eval/exec/compile/Function on hook surfaces.""" match = _DYNAMIC_EVAL.search(content) diff --git a/docs/TRACEABILITY.md b/docs/TRACEABILITY.md index 5b14a3d6..bfc1e5a0 100644 --- a/docs/TRACEABILITY.md +++ b/docs/TRACEABILITY.md @@ -22,7 +22,7 @@ | structural Semgrep-style `pattern:` execution by lightweight engine | built-in scanner | not implemented unless a real structural matcher is added; fixtures are not execution | | GitHub Actions transport-only polling loop (#1087, #938 vertical slice) | owned by PR #1088 / issue #1087; YAML rules and RED precision contracts | mapped-family only; this successor does not ship or close the detector | | Password/database-url/auth-comment precision and test-file context (#1106) | existing `_scan_file` rules `hardcoded-password`, `hardcoded-database-url`, `todo-skip-auth`, `_finding_context` | implemented-branch regression lock | -| Claude plugin marketplace/package supply chain (#1099) | `claude-plugin-floating-git-ref`, `claude-plugin-provider-secret`, `claude-plugin-pipe-to-shell`, `claude-plugin-unsigned-executable-download` (hooks and package.json lifecycle scripts), `claude-plugin-unpinned-package-install`, `claude-plugin-undeclared-executable`, `claude-plugin-symlink-escape`, `claude-plugin-archive-path-traversal`, `claude-plugin-unadmitted-submodule`, `claude-plugin-duplicate-json-member`, `claude-plugin-nonstandard-json-constant`, `claude-plugin-malformed-utf8`, `claude-plugin-inconsistent-normalized-name`, `claude-plugin-vendored-scope-undeclared`, `claude-plugin-conflicting-identity`, `claude-plugin-unbounded-mcp`, `claude-plugin-license-missing`, `claude-plugin-license-mismatch`, `claude-plugin-dynamic-eval`, `claude-plugin-hidden-undeclared-executable`, `claude-plugin-concealed-identity`, `claude-plugin-oversized-package`, `claude-plugin-source-mismatch`, `claude-plugin-github-write-token`, `claude-plugin-docker-socket`, `claude-plugin-browser-profile-access`, `claude-plugin-deceptive-description`, `claude-plugin-secret-to-network`, `claude-plugin-secret-to-prompt`, `claude-plugin-secret-to-mcp`, `claude-plugin-hide-actions-directive` / `claude-plugin-self-modify-directive` / `claude-plugin-goal-escalation-directive`, `claude-plugin-setuid-executable` / `claude-plugin-world-writable-executable`, `claude-plugin-decompression-bomb`, reused #1036 `skill-name-homoglyph-confusable` / `skill-manifest-prompt-injection-payload` / `skill-doc-exfiltration-endpoint-directive` / `skill-placeholder-template-unresolved` on plugin skill/agent/command surfaces, deterministic scan receipt with catalog repository/SHA bind, SARIF 2.1.0 `sarif_sha256` bound to the same finding rule_ids, `policy_provenance` bound to the AppGuardrail release plus exact scan-policy digest, and `sbom_sha256` of a deterministic CycloneDX 1.5 document, `claude-plugin-checksum-mismatch` when a first-party SHA256SUMS or sibling `*.sha256` disagrees with bytes on disk, `claude-plugin-unsigned-checksum` when checksum digest rows have no sibling Cosign/GPG signature file, `claude-plugin-excessive-path-depth` when a materialized file or archive member nests past 32 path components, `claude-plugin-github-merge-command` for hook or manifest `gh pr merge`, `claude-plugin-github-release-command` for `gh release create|upload|delete|edit`, `claude-plugin-kubectl-apply-command` for hook or manifest `kubectl apply`, `claude-plugin-docker-push-command` for `docker push`, `claude-plugin-terraform-apply-command` for `terraform apply`, `claude-plugin-helm-install-command` for `helm install`, `claude-plugin-vercel-deploy-command` for hook or manifest `vercel deploy`, `claude-plugin-fly-deploy-command` for `fly deploy`, `claude-plugin-aws-deploy-command` for hook or manifest `aws cloudformation deploy`, `claude-plugin-gcloud-deploy-command` for `gcloud run deploy`, `claude-plugin-az-deploy-command` for `az webapp deploy`, `claude-plugin-credential-store-access` for host cookie and token stores that are not browser profiles, fail-closed receipt verification | implemented-branch | +| Claude plugin marketplace/package supply chain (#1099) | `claude-plugin-floating-git-ref`, `claude-plugin-provider-secret`, `claude-plugin-pipe-to-shell`, `claude-plugin-unsigned-executable-download` (hooks and package.json lifecycle scripts), `claude-plugin-unpinned-package-install`, `claude-plugin-undeclared-executable`, `claude-plugin-symlink-escape`, `claude-plugin-archive-path-traversal`, `claude-plugin-unadmitted-submodule`, `claude-plugin-duplicate-json-member`, `claude-plugin-nonstandard-json-constant`, `claude-plugin-malformed-utf8`, `claude-plugin-inconsistent-normalized-name`, `claude-plugin-vendored-scope-undeclared`, `claude-plugin-conflicting-identity`, `claude-plugin-unbounded-mcp`, `claude-plugin-license-missing`, `claude-plugin-license-mismatch`, `claude-plugin-dynamic-eval`, `claude-plugin-hidden-undeclared-executable`, `claude-plugin-concealed-identity`, `claude-plugin-oversized-package`, `claude-plugin-source-mismatch`, `claude-plugin-github-write-token`, `claude-plugin-docker-socket`, `claude-plugin-browser-profile-access`, `claude-plugin-deceptive-description`, `claude-plugin-secret-to-network`, `claude-plugin-secret-to-prompt`, `claude-plugin-secret-to-mcp`, `claude-plugin-hide-actions-directive` / `claude-plugin-self-modify-directive` / `claude-plugin-goal-escalation-directive`, `claude-plugin-setuid-executable` / `claude-plugin-world-writable-executable`, `claude-plugin-decompression-bomb`, reused #1036 `skill-name-homoglyph-confusable` / `skill-manifest-prompt-injection-payload` / `skill-doc-exfiltration-endpoint-directive` / `skill-placeholder-template-unresolved` on plugin skill/agent/command surfaces, deterministic scan receipt with catalog repository/SHA bind, SARIF 2.1.0 `sarif_sha256` bound to the same finding rule_ids, `policy_provenance` bound to the AppGuardrail release plus exact scan-policy digest, and `sbom_sha256` of a deterministic CycloneDX 1.5 document, `claude-plugin-checksum-mismatch` when a first-party SHA256SUMS or sibling `*.sha256` disagrees with bytes on disk, `claude-plugin-unsigned-checksum` when checksum digest rows have no sibling Cosign/GPG signature file, `claude-plugin-excessive-path-depth` when a materialized file or archive member nests past 32 path components, `claude-plugin-github-merge-command` for hook or manifest `gh pr merge`, `claude-plugin-github-release-command` for `gh release create|upload|delete|edit`, `claude-plugin-kubectl-apply-command` for hook or manifest `kubectl apply`, `claude-plugin-docker-push-command` for `docker push`, `claude-plugin-terraform-apply-command` for `terraform apply`, `claude-plugin-helm-install-command` for `helm install`, `claude-plugin-vercel-deploy-command` for hook or manifest `vercel deploy`, `claude-plugin-fly-deploy-command` for `fly deploy`, `claude-plugin-aws-deploy-command` for hook or manifest `aws cloudformation deploy`, `claude-plugin-gcloud-deploy-command` for `gcloud run deploy`, `claude-plugin-az-deploy-command` for `az webapp deploy`, `claude-plugin-aws-s3-write-command` for hook or manifest `aws s3 sync`/`cp`, `claude-plugin-az-containerapp-up-command` for `az containerapp up`, `claude-plugin-credential-store-access` for host cookie and token stores that are not browser profiles, fail-closed receipt verification | implemented-branch | | Orphaned GitHub Actions registry identities (#929) | owned by PR #966 / issue #929; live registry DAST | mapped-family only; this successor does not ship or close the detector | | Org security-failure CI tickets without copied vuln evidence | documented non-detectable family | snapshot in `tests/fixtures/cwl-security-issue-inventory.json` | diff --git a/docs/sast-dast-rule-research.md b/docs/sast-dast-rule-research.md index f5f0c68c..9c3f952f 100644 --- a/docs/sast-dast-rule-research.md +++ b/docs/sast-dast-rule-research.md @@ -109,7 +109,10 @@ files being scanned, then applies the union of relevant checks. Examples: `claude-plugin-fly-deploy-command` for ``fly deploy``, `claude-plugin-aws-deploy-command` for ``aws cloudformation deploy``, `claude-plugin-gcloud-deploy-command` for ``gcloud run deploy``, - `claude-plugin-az-deploy-command` for ``az webapp deploy``, and + `claude-plugin-az-deploy-command` for ``az webapp deploy``, + `claude-plugin-aws-s3-write-command` for ``aws s3 sync``/``cp``, + `claude-plugin-az-containerapp-up-command` for ``az containerapp up``, + and `claude-plugin-credential-store-access` for host ``~/.netrc``, ``~/.aws/credentials``, GitHub CLI hosts, Docker auth, cookie jars, and SSH private keys. Chrome/Firefox profile stores stay diff --git a/tests/test_claude_plugin_cloud_deploy.py b/tests/test_claude_plugin_cloud_deploy.py index 01fb690d..e1b948f7 100644 --- a/tests/test_claude_plugin_cloud_deploy.py +++ b/tests/test_claude_plugin_cloud_deploy.py @@ -196,6 +196,7 @@ def test_plugin_manifest_az_webapp_deploy_fails_admission(tmp_path: Path) -> Non (root / ".claude-plugin" / "plugin.json").read_text(encoding="utf-8") ) manifest["hooks"] = { + "PreToolUse": [{"command": "hooks/session.sh"}], "PostToolUse": [{"command": "az webapp deploy --name app"}], } _write_json(root / ".claude-plugin" / "plugin.json", manifest) @@ -225,6 +226,7 @@ def test_manifest_cloud_prose_and_reporting_commands_are_not_this_class( manifest = json.loads(manifest_path.read_text(encoding="utf-8")) manifest["description"] = "operators may later run gcloud run deploy" manifest["hooks"] = { + "PreToolUse": [{"command": "hooks/session.sh"}], "PostToolUse": [ {"command": 'echo "aws cloudformation deploy"'}, {"command": "printf '%s\\n' 'az webapp deploy'"}, @@ -246,6 +248,7 @@ def test_manifest_reporting_command_does_not_hide_later_cloud_deploy( manifest_path = root / ".claude-plugin" / "plugin.json" manifest = json.loads(manifest_path.read_text(encoding="utf-8")) manifest["hooks"] = { + "PreToolUse": [{"command": "hooks/session.sh"}], "PostToolUse": [ {"command": 'echo "gcloud run deploy"'}, {"command": "aws cloudformation deploy --stack-name app"}, diff --git a/tests/test_claude_plugin_hosted_deploy.py b/tests/test_claude_plugin_hosted_deploy.py index 4464cca0..447e0efd 100644 --- a/tests/test_claude_plugin_hosted_deploy.py +++ b/tests/test_claude_plugin_hosted_deploy.py @@ -251,6 +251,7 @@ def test_plugin_manifest_fly_deploy_fails_admission(tmp_path: Path) -> None: (root / ".claude-plugin" / "plugin.json").read_text(encoding="utf-8") ) manifest["hooks"] = { + "PreToolUse": [{"command": "hooks/session.sh"}], "PostToolUse": [{"command": "fly deploy --now"}], } _write_json(root / ".claude-plugin" / "plugin.json", manifest) @@ -345,6 +346,7 @@ def test_manifest_reporting_commands_and_description_are_not_this_class( manifest = json.loads(manifest_path.read_text(encoding="utf-8")) manifest["description"] = "operators may later run vercel deploy" manifest["hooks"] = { + "PreToolUse": [{"command": "hooks/session.sh"}], "PostToolUse": [ {"command": 'echo "fly deploy"'}, {"command": "printf '%s\\n' 'vercel deploy'"}, @@ -368,6 +370,7 @@ def test_manifest_reporting_command_does_not_hide_later_deploy( manifest_path = root / ".claude-plugin" / "plugin.json" manifest = json.loads(manifest_path.read_text(encoding="utf-8")) manifest["hooks"] = { + "PreToolUse": [{"command": "hooks/session.sh"}], "PostToolUse": [ {"command": 'echo "fly deploy"'}, {"command": "vercel deploy --prod"}, diff --git a/tests/test_claude_plugin_object_store.py b/tests/test_claude_plugin_object_store.py index 3e9dda14..1abce52e 100644 --- a/tests/test_claude_plugin_object_store.py +++ b/tests/test_claude_plugin_object_store.py @@ -161,6 +161,7 @@ def test_plugin_manifest_containerapp_up_fails_admission(tmp_path: Path) -> None (root / ".claude-plugin" / "plugin.json").read_text(encoding="utf-8") ) manifest["hooks"] = { + "PreToolUse": [{"command": "hooks/session.sh"}], "PostToolUse": [{"command": "az containerapp up --name app"}], } _write_json(root / ".claude-plugin" / "plugin.json", manifest) @@ -177,6 +178,7 @@ def test_manifest_prose_and_reporting_are_not_this_class(tmp_path: Path) -> None manifest = json.loads(manifest_path.read_text(encoding="utf-8")) manifest["description"] = "operators may later run aws s3 sync" manifest["hooks"] = { + "PreToolUse": [{"command": "hooks/session.sh"}], "PostToolUse": [{"command": 'echo "az containerapp up"'}], } _write_json(manifest_path, manifest) From f480ae0e9805e49f5007bbd8abd5c5cda3c9c487 Mon Sep 17 00:00:00 2001 From: Seongho Bae Date: Tue, 8 Sep 2026 14:10:18 +0900 Subject: [PATCH 03/12] test(scanner): distinguish s3 downloads from writes RED --- tests/test_claude_plugin_object_store.py | 27 +++++++++++++++++++++--- 1 file changed, 24 insertions(+), 3 deletions(-) diff --git a/tests/test_claude_plugin_object_store.py b/tests/test_claude_plugin_object_store.py index 1abce52e..f809921f 100644 --- a/tests/test_claude_plugin_object_store.py +++ b/tests/test_claude_plugin_object_store.py @@ -1,4 +1,4 @@ -"""Hook aws s3 writes and az containerapp up fail closed; ls stays inventory.""" +"""Hook aws s3 writes and az containerapp up fail closed; reads stay inventory.""" from __future__ import annotations @@ -58,7 +58,7 @@ def _hits(root: Path, rule_id: str): def test_hook_aws_s3_sync_fails_admission(tmp_path: Path) -> None: - """``aws s3 sync`` on a hook is object-store write authority.""" + """``aws s3 sync`` to S3 is object-store write authority.""" root = _licensed_plugin(tmp_path, "#!/bin/sh\naws s3 sync ./dist s3://bucket/app\n") hits = _hits(root, _S3_RULE) receipt = build_claude_plugin_scan_receipt(root) @@ -74,12 +74,33 @@ def test_hook_aws_s3_sync_fails_admission(tmp_path: Path) -> None: def test_hook_aws_s3_cp_fails_admission() -> None: - """``aws s3 cp`` is the same object-store write class.""" + """``aws s3 cp`` to S3 is object-store write authority.""" body = "#!/bin/sh\naws s3 cp artifact.tgz s3://bucket/app.tgz\n" hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) assert any(hit.rule_id == _S3_RULE and hit.snippet == "aws s3 cp" for hit in hits) +def test_hook_aws_s3_cp_download_stays_inventory() -> None: + """A provable S3-to-local copy is read authority, not this write class.""" + body = "#!/bin/sh\naws s3 cp s3://bucket/app.tgz ./app.tgz\n" + hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) + assert not any(hit.rule_id == _S3_RULE for hit in hits) + + +def test_hook_aws_s3_sync_download_stays_inventory() -> None: + """A provable S3-to-local sync is read authority, not this write class.""" + body = "#!/bin/sh\naws s3 sync s3://bucket/app ./app\n" + hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) + assert not any(hit.rule_id == _S3_RULE for hit in hits) + + +def test_hook_aws_s3_to_s3_copy_still_fails_admission() -> None: + """S3-to-S3 copy still writes the destination bucket.""" + body = "#!/bin/sh\naws s3 cp s3://source/app.tgz s3://target/app.tgz\n" + hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) + assert any(hit.rule_id == _S3_RULE for hit in hits) + + def test_hook_az_containerapp_up_fails_admission(tmp_path: Path) -> None: """``az containerapp up`` on a hook is Azure write authority.""" root = _licensed_plugin(tmp_path, "#!/bin/sh\naz containerapp up --name app\n") From 51d1da4ba25d88c5036c42d519475d8cdf871e8a Mon Sep 17 00:00:00 2001 From: Seongho Bae Date: Tue, 8 Sep 2026 14:20:53 +0900 Subject: [PATCH 04/12] fix(scanner): distinguish literal S3 downloads from writes --- appguardrail_core/claude_plugin_detector.py | 29 +++++++++++++++++++++ 1 file changed, 29 insertions(+) diff --git a/appguardrail_core/claude_plugin_detector.py b/appguardrail_core/claude_plugin_detector.py index f9edabeb..546ec5c6 100644 --- a/appguardrail_core/claude_plugin_detector.py +++ b/appguardrail_core/claude_plugin_detector.py @@ -2011,6 +2011,33 @@ def _az_deploy_command_hits( return () +def _is_literal_s3_download( + content: str, match: re.Match[str] +) -> bool: + """Return whether one direct S3 command provably reads to a local path.""" + line_start = content.rfind("\n", 0, match.start()) + 1 + line_end = content.find("\n", match.start()) + if line_end < 0: + line_end = len(content) + line = content[line_start:line_end] + relative = match.start() - line_start + command_end = match.end() - line_start + for start, end in _iter_unquoted_segment_bounds(line): + if start <= relative < end: + operands = line[command_end:end].split() + return ( + len(operands) == 2 + and operands[0].lower().startswith("s3://") + and not operands[1].lower().startswith("s3://") + and not operands[1].startswith("-") + and all( + re.fullmatch(r"[A-Za-z0-9._~:/+-]+", operand) + for operand in operands + ) + ) + return False + + def _aws_s3_write_command_hits( content: str, *, manifest: bool = False ) -> tuple[PluginHit, ...]: @@ -2028,6 +2055,8 @@ def _aws_s3_write_command_hits( match = _executable_command_match(source, _AWS_S3_WRITE_COMMAND) if match is None: continue + if _is_literal_s3_download(source, match): + continue verb = match.group("verb").lower() return ( PluginHit( From 44761c28d8748f63eca052ec6145be1111bc4d68 Mon Sep 17 00:00:00 2001 From: Seongho Bae Date: Tue, 8 Sep 2026 14:21:30 +0900 Subject: [PATCH 05/12] test(scanner): reproduce download-before-upload false negative --- tests/test_claude_plugin_object_store.py | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/tests/test_claude_plugin_object_store.py b/tests/test_claude_plugin_object_store.py index f809921f..6c9e2bae 100644 --- a/tests/test_claude_plugin_object_store.py +++ b/tests/test_claude_plugin_object_store.py @@ -219,3 +219,15 @@ def test_cloudformation_deploy_stays_the_aws_deploy_class() -> None: rule_ids = {hit.rule_id for hit in hits} assert _AWS_DEPLOY_RULE in rule_ids assert _THIS_CLASS.isdisjoint(rule_ids) + + +def test_download_before_later_upload_still_fails_admission() -> None: + """A safe first command cannot hide a later S3 destination write.""" + body = ( + "#!/bin/sh\n" + "aws s3 cp s3://bucket/input.tgz ./input.tgz && " + "aws s3 cp ./output.tgz s3://bucket/output.tgz\n" + ) + hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) + assert any(hit.rule_id == _S3_RULE for hit in hits) + From 3073ffdbbddbaa30d363f366d7dc5deea250f939 Mon Sep 17 00:00:00 2001 From: Seongho Bae Date: Tue, 8 Sep 2026 14:22:21 +0900 Subject: [PATCH 06/12] fix(scanner): scan past safe S3 downloads --- appguardrail_core/claude_plugin_detector.py | 43 +++++++++------------ 1 file changed, 19 insertions(+), 24 deletions(-) diff --git a/appguardrail_core/claude_plugin_detector.py b/appguardrail_core/claude_plugin_detector.py index 546ec5c6..46b3a152 100644 --- a/appguardrail_core/claude_plugin_detector.py +++ b/appguardrail_core/claude_plugin_detector.py @@ -2041,31 +2041,26 @@ def _is_literal_s3_download( def _aws_s3_write_command_hits( content: str, *, manifest: bool = False ) -> tuple[PluginHit, ...]: - """Return ``aws s3 sync``/``cp`` findings with a command label, not URIs. - - Args: - content: Hook or manifest text. - manifest: When true, only structural command values are scanned. - - Returns: - One hit for executable ``aws s3 sync`` or ``aws s3 cp``. - ``aws s3 ls``, comments, and echo lookalikes are not this class. - """ + """Return S3-destination write findings without copying operand URIs.""" for source, first_line in _hosted_command_sources(content, manifest=manifest): - match = _executable_command_match(source, _AWS_S3_WRITE_COMMAND) - if match is None: - continue - if _is_literal_s3_download(source, match): - continue - verb = match.group("verb").lower() - return ( - PluginHit( - rule_id="claude-plugin-aws-s3-write-command", - line=first_line + source[: match.start()].count("\n"), - snippet="aws s3 " + verb, - message=CLAUDE_PLUGIN_AWS_S3_WRITE_COMMAND_MESSAGE, - ), - ) + search_from = 0 + while search_from < len(source): + fragment = source[search_from:] + match = _executable_command_match(fragment, _AWS_S3_WRITE_COMMAND) + if match is None: + break + absolute_start = search_from + match.start() + if not _is_literal_s3_download(fragment, match): + verb = match.group("verb").lower() + return ( + PluginHit( + rule_id="claude-plugin-aws-s3-write-command", + line=first_line + source[:absolute_start].count("\n"), + snippet="aws s3 " + verb, + message=CLAUDE_PLUGIN_AWS_S3_WRITE_COMMAND_MESSAGE, + ), + ) + search_from += match.end() return () From 7f30f705e36cc127b4825948ec00481b5de12720 Mon Sep 17 00:00:00 2001 From: Seongho Bae Date: Tue, 8 Sep 2026 16:12:08 +0900 Subject: [PATCH 07/12] fix(scanner): inherit quoted command-context repair --- appguardrail_core/claude_plugin_detector.py | 84 ++++++++++++++++++--- 1 file changed, 74 insertions(+), 10 deletions(-) diff --git a/appguardrail_core/claude_plugin_detector.py b/appguardrail_core/claude_plugin_detector.py index 46b3a152..b0877d53 100644 --- a/appguardrail_core/claude_plugin_detector.py +++ b/appguardrail_core/claude_plugin_detector.py @@ -1860,14 +1860,73 @@ def _hosted_command_sources( return ((content, 1),) +def _shell_command_context_start(line: str, offset: int) -> int | None: + """Return the executable shell-frame start containing ``offset``. + + Args: + line: One hook or manifest command line. + offset: Zero-based match offset within ``line``. + + Returns: + The start of the root, ``$(...)``, or backtick command frame. + ``None`` means the offset is inert single- or double-quoted prose. + """ + frames: list[tuple[str, int, str, int]] = [("", 0, "", 0)] + escaped = False + index = 0 + while index < offset: + frame_end, frame_start, quote, depth = frames[-1] + char = line[index] + if escaped: + escaped = False + index += 1 + continue + if char == "\\" and quote != "'": + escaped = True + index += 1 + continue + if char == "'" and quote != '"': + frames[-1] = (frame_end, frame_start, "" if quote == "'" else "'", depth) + index += 1 + continue + if char == '"' and quote != "'": + frames[-1] = (frame_end, frame_start, "" if quote == '"' else '"', depth) + index += 1 + continue + if quote != "'" and line[index : index + 2] == "$(": + frames.append((")", index + 2, "", 1)) + index += 2 + continue + if quote != "'" and char == "`": + if frame_end == "`": + frames.pop() + else: + frames.append(("`", index + 1, "", 0)) + index += 1 + continue + if quote: + index += 1 + continue + if frame_end == ")" and char == "(": + frames[-1] = (frame_end, frame_start, quote, depth + 1) + elif frame_end == ")" and char == ")": + if depth == 1: + frames.pop() + else: + frames[-1] = (frame_end, frame_start, quote, depth - 1) + index += 1 + _frame_end, frame_start, quote, _depth = frames[-1] + return None if quote else frame_start + + def _executable_command_match( content: str, pattern: re.Pattern[str] ) -> re.Match[str] | None: """Return the first regex match that is an executable command context. - Unquoted ``#`` comments and ``echo``/``printf``/``print`` segments are - not executable. Manifest JSON command strings remain searchable - because they are not reporting builtins. + Unquoted ``#`` comments, quoted prose, and + ``echo``/``printf``/``print`` segments are not executable. Direct + commands inside ``$(...)`` or backticks remain executable. Args: content: Hook or manifest text. @@ -1885,17 +1944,22 @@ def _executable_command_match( line_end = len(content) line = content[line_start:line_end] relative = match.start() - line_start - comment_at = _unquoted_hash_index(line) - if comment_at is not None and relative >= comment_at: + context_start = _shell_command_context_start(line, relative) + if context_start is None: continue - for start, end in _iter_unquoted_segment_bounds(line): - if start <= relative < end: - if not _is_reporting_builtin_segment(line[start:end]): + context = line[context_start:] + context_relative = relative - context_start + comment_at = _unquoted_hash_index(context) + if comment_at is not None and context_relative >= comment_at: + continue + for segment_start, segment_end in _iter_unquoted_segment_bounds(context): + if segment_start <= context_relative < segment_end: + if not _is_reporting_builtin_segment( + context[segment_start:segment_end] + ): return match break return None - - def _vercel_deploy_command_hits( content: str, *, manifest: bool = False ) -> tuple[PluginHit, ...]: From f130c2cdc023827cddd2364039753d84ad801cd9 Mon Sep 17 00:00:00 2001 From: Seongho Bae Date: Tue, 8 Sep 2026 16:12:09 +0900 Subject: [PATCH 08/12] test(scanner): inherit quoted command-context regressions --- tests/test_claude_plugin_terraform_helm.py | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+) diff --git a/tests/test_claude_plugin_terraform_helm.py b/tests/test_claude_plugin_terraform_helm.py index 4a95efdf..d413aa2c 100644 --- a/tests/test_claude_plugin_terraform_helm.py +++ b/tests/test_claude_plugin_terraform_helm.py @@ -254,3 +254,25 @@ def test_later_executable_command_after_reporting_segment_still_fails() -> None: hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) assert any(hit.rule_id in _THIS_CLASS for hit in hits) +def test_quoted_shell_prose_is_not_executable() -> None: + """Quoted command names and reporting substitutions are inert prose.""" + bodies = ( + '#!/bin/sh\nmessage="terraform apply -auto-approve"\n', + '#!/bin/sh\nif [ "$mode" = "helm install app chart/" ]; then echo safe; fi\n', + "#!/bin/sh\nmessage='helm install app chart/'\n", + '#!/bin/sh\nresult="$(echo \'terraform apply -auto-approve\')"\n', + ) + for body in bodies: + hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) + assert _THIS_CLASS.isdisjoint(hit.rule_id for hit in hits) + + +def test_command_substitution_remains_executable() -> None: + """Direct commands in modern and legacy substitutions remain executable.""" + bodies = ( + '#!/bin/sh\nresult="$(terraform apply -auto-approve)"\n', + "#!/bin/sh\nresult=`helm install app chart/`\n", + ) + for body in bodies: + hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) + assert any(hit.rule_id in _THIS_CLASS for hit in hits) From 097b6dddfdba7ea2a8cd069942f6371c269e0e6a Mon Sep 17 00:00:00 2001 From: Seongho Bae Date: Tue, 8 Sep 2026 16:37:18 +0900 Subject: [PATCH 09/12] test(scanner): inherit assignment command boundary --- tests/test_claude_plugin_terraform_helm.py | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+) diff --git a/tests/test_claude_plugin_terraform_helm.py b/tests/test_claude_plugin_terraform_helm.py index d413aa2c..b80b6877 100644 --- a/tests/test_claude_plugin_terraform_helm.py +++ b/tests/test_claude_plugin_terraform_helm.py @@ -276,3 +276,25 @@ def test_command_substitution_remains_executable() -> None: for body in bodies: hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) assert any(hit.rule_id in _THIS_CLASS for hit in hits) + + +def test_assignment_values_are_not_executable_commands() -> None: + """An unquoted assignment value cannot turn its following word into the CLI.""" + bodies = ( + "#!/bin/sh\nmessage=terraform apply -auto-approve\n", + "#!/bin/sh\ncommand=helm install app chart/\n", + ) + for body in bodies: + hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) + assert _THIS_CLASS.isdisjoint(hit.rule_id for hit in hits) + + +def test_environment_assignment_before_real_command_still_fails() -> None: + """Environment assignments do not hide a later executable deployment CLI.""" + bodies = ( + "#!/bin/sh\nTF_IN_AUTOMATION=1 terraform apply -auto-approve\n", + "#!/bin/sh\nHELM_NAMESPACE=prod helm install app chart/\n", + ) + for body in bodies: + hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) + assert any(hit.rule_id in _THIS_CLASS for hit in hits) From 5e8b1987daf52466aeb1c3877bcc95cfd59c046f Mon Sep 17 00:00:00 2001 From: Seongho Bae Date: Tue, 8 Sep 2026 16:37:19 +0900 Subject: [PATCH 10/12] fix(scanner): inherit assignment command boundary --- appguardrail_core/claude_plugin_detector.py | 28 ++++++++++++++++++--- 1 file changed, 24 insertions(+), 4 deletions(-) diff --git a/appguardrail_core/claude_plugin_detector.py b/appguardrail_core/claude_plugin_detector.py index b0877d53..cc0bcf82 100644 --- a/appguardrail_core/claude_plugin_detector.py +++ b/appguardrail_core/claude_plugin_detector.py @@ -1919,12 +1919,28 @@ def _shell_command_context_start(line: str, offset: int) -> int | None: return None if quote else frame_start -def _executable_command_match( - content: str, pattern: re.Pattern[str] +def _match_starts_in_shell_assignment_value(segment: str, offset: int) -> bool: + """Return whether ``offset`` starts inside an unquoted assignment word. + + Args: + segment: One shell command segment. + offset: Zero-based match offset within ``segment``. + + Returns: + True when the current shell word before ``offset`` contains ``=``. + An assignment followed by whitespace and a real command returns False. + """ + prefix = segment[:offset] + if not prefix or prefix[-1].isspace(): + return False + return "=" in prefix.rsplit(maxsplit=1)[-1] + + +def _executable_command_match( content: str, pattern: re.Pattern[str] ) -> re.Match[str] | None: """Return the first regex match that is an executable command context. - Unquoted ``#`` comments, quoted prose, and + Unquoted ``#`` comments, quoted prose, shell assignment values, and ``echo``/``printf``/``print`` segments are not executable. Direct commands inside ``$(...)`` or backticks remain executable. @@ -1954,8 +1970,12 @@ def _executable_command_match( continue for segment_start, segment_end in _iter_unquoted_segment_bounds(context): if segment_start <= context_relative < segment_end: + segment = context[segment_start:segment_end] + segment_relative = context_relative - segment_start if not _is_reporting_builtin_segment( - context[segment_start:segment_end] + segment + ) and not _match_starts_in_shell_assignment_value( + segment, segment_relative ): return match break From f5dd35dccb205054c310724e3590d4e47614a10a Mon Sep 17 00:00:00 2001 From: Seongho Bae Date: Tue, 8 Sep 2026 17:14:39 +0900 Subject: [PATCH 11/12] fix(scanner): carry heredoc command boundary --- appguardrail_core/claude_plugin_detector.py | 57 ++++++++++++++++++++- 1 file changed, 55 insertions(+), 2 deletions(-) diff --git a/appguardrail_core/claude_plugin_detector.py b/appguardrail_core/claude_plugin_detector.py index cc0bcf82..4f29e0cd 100644 --- a/appguardrail_core/claude_plugin_detector.py +++ b/appguardrail_core/claude_plugin_detector.py @@ -450,6 +450,11 @@ ) _REPORTING_BUILTINS: Final = frozenset({"echo", "printf", "print"}) _FIRST_SHELL_TOKEN = re.compile(r"\s*([A-Za-z0-9_./+-]+)") +_LITERAL_HEREDOC_OPEN = re.compile( + r"<<(?P-)?[ \t]*(?P['\"]?)" + r"(?P[A-Za-z_][A-Za-z0-9_]*)(?P=quote)" + r"(?=$|[ \t;&|()<>])" +) _DOCKER_SOCKET = re.compile( r"(?:/var/run/docker\.sock|unix://\S*docker\.sock)", re.IGNORECASE, @@ -1919,6 +1924,50 @@ def _shell_command_context_start(line: str, offset: int) -> int | None: return None if quote else frame_start +def _literal_heredoc_payload_spans(content: str) -> tuple[tuple[int, int], ...]: + """Return closed literal here-document payload spans. + + Args: + content: One hook or structural manifest command string. + + Returns: + Inclusive-start exclusive-end spans for payloads with one confidently + parsed identifier delimiter on the opener line. Quoted delimiters and + tab-stripping forms are supported. Ambiguous or unclosed forms stay + executable for fail-closed analysis. + """ + spans: list[tuple[int, int]] = [] + active: tuple[str, bool, int] | None = None + offset = 0 + for raw_line in content.splitlines(keepends=True): + line = raw_line.rstrip("\r\n") + if active is not None: + delimiter, strip_tabs, payload_start = active + candidate = line.lstrip("\t") if strip_tabs else line + if candidate == delimiter: + spans.append((payload_start, offset)) + active = None + offset += len(raw_line) + continue + + comment_at = _unquoted_hash_index(line) + openers = tuple( + match + for match in _LITERAL_HEREDOC_OPEN.finditer(line) + if (comment_at is None or match.start() < comment_at) + and _shell_command_context_start(line, match.start()) is not None + ) + if len(openers) == 1: + opener = openers[0] + active = ( + opener.group("delimiter"), + opener.group("strip") is not None, + offset + len(raw_line), + ) + offset += len(raw_line) + return tuple(spans) + + def _match_starts_in_shell_assignment_value(segment: str, offset: int) -> bool: """Return whether ``offset`` starts inside an unquoted assignment word. @@ -1940,8 +1989,9 @@ def _executable_command_match( content: str, pattern: re.Pattern[str] ) -> re.Match[str] | None: """Return the first regex match that is an executable command context. - Unquoted ``#`` comments, quoted prose, shell assignment values, and - ``echo``/``printf``/``print`` segments are not executable. Direct + Unquoted ``#`` comments, quoted prose, closed literal here-document + payloads, shell assignment values, and ``echo``/``printf``/``print`` + segments are not executable. Direct commands inside ``$(...)`` or backticks remain executable. Args: @@ -1953,7 +2003,10 @@ def _executable_command_match( content: str, pattern: re.Pattern[str] """ if not content: return None + inert_payloads = _literal_heredoc_payload_spans(content) for match in pattern.finditer(content): + if any(start <= match.start() < end for start, end in inert_payloads): + continue line_start = content.rfind("\n", 0, match.start()) + 1 line_end = content.find("\n", match.start()) if line_end < 0: From e527fd2404db812c7f51ed557c9b9109191fc5df Mon Sep 17 00:00:00 2001 From: Seongho Bae Date: Tue, 8 Sep 2026 17:14:40 +0900 Subject: [PATCH 12/12] test(scanner): carry heredoc command regressions --- tests/test_claude_plugin_terraform_helm.py | 37 ++++++++++++++++++++++ 1 file changed, 37 insertions(+) diff --git a/tests/test_claude_plugin_terraform_helm.py b/tests/test_claude_plugin_terraform_helm.py index b80b6877..fe3bc41e 100644 --- a/tests/test_claude_plugin_terraform_helm.py +++ b/tests/test_claude_plugin_terraform_helm.py @@ -212,6 +212,7 @@ def test_kubectl_apply_without_terraform_stays_the_kubectl_class() -> None: assert _KUBECTL_RULE in rule_ids assert _THIS_CLASS.isdisjoint(rule_ids) + def test_hook_comments_and_reporting_builtins_are_not_commands() -> None: """Comments and reporting builtins do not execute terraform or Helm.""" bodies = ( @@ -278,6 +279,7 @@ def test_command_substitution_remains_executable() -> None: assert any(hit.rule_id in _THIS_CLASS for hit in hits) + def test_assignment_values_are_not_executable_commands() -> None: """An unquoted assignment value cannot turn its following word into the CLI.""" bodies = ( @@ -298,3 +300,38 @@ def test_environment_assignment_before_real_command_still_fails() -> None: for body in bodies: hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) assert any(hit.rule_id in _THIS_CLASS for hit in hits) + + +def test_here_document_payload_is_not_an_executable_command() -> None: + """Literal here-document payload is data, even when it names deployment CLIs.""" + bodies = ( + "#!/bin/sh\ncat <<'EOF'\nterraform apply -auto-approve\nEOF\n", + "#!/bin/sh\ncat <<-EOF\n\thelm install app chart/\n\tEOF\n", + ) + for body in bodies: + hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) + assert _THIS_CLASS.isdisjoint(hit.rule_id for hit in hits) + + +def test_command_after_here_document_still_fails() -> None: + """An inert payload cannot hide a later executable deployment command.""" + body = ( + "#!/bin/sh\ncat <<'EOF'\nterraform apply -auto-approve\nEOF\n" + "helm install app chart/\n" + ) + hits = inspect_claude_plugin_file("session.sh", "hooks/session.sh", body) + rule_ids = {hit.rule_id for hit in hits} + assert _TERRAFORM_RULE not in rule_ids + assert _HELM_RULE in rule_ids + + +def test_heredoc_opener_lookalikes_do_not_hide_real_commands() -> None: + """Quoted or commented opener text cannot suppress a later real command.""" + bodies = ( + '#!/bin/sh\necho "<