docs: working with non-Anthropic models - #11
Open
schenksj wants to merge 3 commits into
Open
Conversation
Add docs/working-with-non-anthropic-models.md covering the ~/.local/bin/openrouter_claude wrapper (OpenRouter env overrides, single-model tier pinning, read-only-by-default because Bash is omitted, Agent as the load-bearing tool), a quick start, and the currently recommended models (Opus 4.8 native, GLM-5.2 via OpenRouter) with the not-recommended set and the reasons (Kimi noisy/slow, Nemotron incomplete, Gemma failed). Documents the criteria a model must meet, grounded in the cross-model eval: baseline (1M+ context, instruction-following, reasoning, code understanding) plus the eval-derived discriminators (agentic subagent orchestration, long-horizon stamina, output consistency & valid citations, trust-boundary judgment & false-positive discipline). Notes OpenAI frontier models are planned under OpenAI's trusted cyber program (separate access path) and not yet tested. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fold in two new cross-model evaluation runs: - Kimi K3 (moonshotai/kimi-k3): now recommended on a single run — led the corpus on true-positive count and cost-effectiveness with strong precision, and uniquely discovered two exploitable High-severity vulnerabilities no other model found (a dot-segment token path-prefix bypass and an ambient- credential scp-form clone). Caveats noted: single run, ~5.5 h, and its severe coverage is credit for vulns it discovered. Distinguished from the older, superseded Kimi k2.7-code. - Qwen 3.7 Max (qwen/qwen3.7-max): not recommended — strong recon but the hunt stage mass-dismissed nearly every real finding (1 TP + 1 FP vs 37 known vulns, missing all five Highs) despite 40 subagents and 1,100+ tool calls; also self-mislabeled its model and declared the codebase "well-hardened." Updates the TL;DR accordingly. Full data in eval_rep/EVAL_REPORT.md. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds Opus 5 (11th evaluated run) to the recommended-models table and TL;DR. 32 adjudicated-real findings at 0.865 precision — more than double any other model — but $110.07/run and third place when scored only on the vulnerabilities every model was measured against. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds
docs/working-with-non-anthropic-models.md— how to run the VulnHunter skills on non-Anthropic models via OpenRouter, and which models actually work, grounded in a cross-model evaluation.Contents
Bashfor a static (not merely read-only) run, with the containment rationale for untrusted third-party/open-weight models.Bashis blocked, and the invocation that opts into a non-recommended model past the model gate.🤖 Generated with Claude Code