Skip to content

Security: ryktara/Foundry

Security

SECURITY.md

Security

Reporting

Email security@alsadq.ae with the affected file, the version (.claude-plugin/plugin.json) and steps to reproduce. Please do not open a public issue for an unpatched vulnerability. You will get an acknowledgement within 5 working days.

Scope

Foundry is a Claude Code plugin: Python scripts, hooks, skills and domain packs that run inside a logged-in Claude Code session. It reads no API key and has no headless mode. Generated apps are the user's; their security posture is checked by the pipeline's threat model, security-review subagent and compliance evidence, not guaranteed by this repository.

Agentic controls (OWASP Top 10 for Agentic Applications 2026)

data/security-controls.csv carries the SEC-AGT family; the pipeline applies it to itself:

id risk what Foundry does
SEC-AGT-01 goal hijack repository content, tickets and web pages are data, never instructions
SEC-AGT-02 tool misuse builders may run only the DoD commands named by the stack and the ticket; hooks guard shell commands
SEC-AGT-03 identity and privilege abuse no production credentials in the build environment; wizards hand secrets to a human
SEC-AGT-04 supply chain skills, packs and MCP servers are validated (foundry.py validate) before use
SEC-AGT-05 unexpected code execution generated code runs only inside DoD commands, in CI or a sandbox
SEC-AGT-06 memory and context poisoning .foundry/ artifacts pass gates before reuse; no untrusted file becomes a skill
SEC-AGT-07 insecure inter-agent communication fixed JSON return contracts; subagent output is data
SEC-AGT-08 cascading failures retry once per gate, a 120-tool-call builder budget, then stop
SEC-AGT-09 human trust exploitation no gate is reported passed without its output; DoD tails are recorded
SEC-AGT-10 rogue agents file-scope limits per ticket and a blast-radius check on every edit
SEC-AGT-11 insecure output handling generated SQL, shell and HTML go through security-review before merge
SEC-AGT-12 excessive agency money-moving paths need a human-confirmed wizard step before live keys

Regulated packs additionally block release until a named person runs release confirm --by <name>.

There aren't any published security advisories