{"id":23153,"date":"2009-05-21T01:01:11","date_gmt":"2009-05-21T01:01:11","guid":{"rendered":"https:\/\/scannn.com\/evilsocket-audit-an-8-stage-vulnerability-discovery-agent-%c2%b7-github\/"},"modified":"2009-05-21T01:01:11","modified_gmt":"2009-05-21T01:01:11","slug":"evilsocket-audit-an-8-stage-vulnerability-discovery-agent-%c2%b7-github","status":"publish","type":"post","link":"https:\/\/scannn.com\/lv\/evilsocket-audit-an-8-stage-vulnerability-discovery-agent-%c2%b7-github\/","title":{"rendered":"evilsocket\/audit: An 8-stage vulnerability-discovery agent. \u00b7 GitHub"},"content":{"rendered":"\n
\n

An 8-stage vulnerability-discovery agent, driven by your Claude Pro \/ Max
\nsubscription<\/strong> through the official Claude Code Agent SDK. Many narrow agents,
\ndeliberate disagreement, and an explicit reachability gate.<\/p>\n

MIT-licensed. No API key needed if you already use claude login<\/code>.<\/p>\n

This project is a from-scratch reimplementation of the pipeline described in
\nCloudflare’s Project Glasswing<\/a>
\npost, which tested Anthropic’s Mythos preview LLM against Cloudflare’s own
\ncodebase. The blog argues that real-world vulnerability discovery does not<\/strong>
\ncome from asking one big model “find bugs here” \u2014 it comes from:<\/p>\n

    \n
  1. Many narrow agents<\/strong> working in parallel on tightly-scoped questions
    \n(“Look for command injection in this specific function, with this trust
    \nboundary above it”) rather than one exhaustive agent.<\/li>\n
  2. Deliberate disagreement<\/strong> \u2014 a second agent, on a different model, that
    \ntries to disprove<\/em> the first agent’s findings.<\/li>\n
  3. A reachability trace<\/strong> as the gating step \u2014 most “is this code buggy?”
    \nfindings are noise unless an attacker-controlled input can actually reach
    \nthe sink from outside the system.<\/li>\n
  4. A feedback loop<\/strong> so reachable bugs in one place automatically seed
    \nhunts for the same pattern elsewhere.<\/li>\n<\/ol>\n

    This repo packages that pipeline into a runnable agent. The Cloudflare post
    \nshowed the architecture; this codebase ships the prompts, schemas, state
    \nstore, and orchestrator.<\/p>\n

    <\/a><\/p>\n

    Diagram from Cloudflare’s Project Glasswing<\/a> post, reproduced here for reference.<\/sub><\/p>\n

    <\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n
    #<\/th>\nStage<\/th>\nDefault model<\/th>\nPurpose<\/th>\n<\/tr>\n<\/thead>\n
    1<\/td>\nRecon<\/td>\nOpus 4.7<\/td>\nMap the repo, emit narrowly-scoped Hunt tasks<\/td>\n<\/tr>\n
    2<\/td>\nHunt<\/td>\nSonnet 4.6<\/td>\nOne attack class per agent; compile\/run PoCs<\/td>\n<\/tr>\n
    3<\/td>\nValidate<\/td>\nOpus 4.7<\/td>\nAdversarial re-read; tries to disprove<\/strong> (different model from Hunt)<\/td>\n<\/tr>\n
    4<\/td>\nGapfill<\/td>\nSonnet 4.6<\/td>\nRe-queue under-covered areas<\/td>\n<\/tr>\n
    5<\/td>\nDedupe<\/td>\nSonnet 4.6<\/td>\nCluster findings by root cause<\/td>\n<\/tr>\n
    6<\/td>\nTrace<\/td>\nOpus 4.7<\/td>\nProve attacker-controlled input reaches the sink<\/td>\n<\/tr>\n
    7<\/td>\nFeedback<\/td>\nSonnet 4.6<\/td>\nTurn reachable traces into new Hunt tasks<\/td>\n<\/tr>\n
    8<\/td>\nReport<\/td>\nSonnet 4.6<\/td>\nSchema-validated structured report<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    <\/markdown-accessiblity-table><\/p>\n

    Each stage is one markdown prompt in prompts\/<\/code> + one JSON Schema in
    \nschemas\/<\/code>. The orchestrator passes the schema into the system prompt so
    \nevery output is shape-stable on the first try.<\/p>\n

    \n
    #<\/span> 1. Install<\/span>\npython -m venv .venv &&<\/span> source<\/span> .venv\/bin\/activate\npip install -e .<\/span>\n\n#<\/span> 2. Auth (pick one)<\/span>\n#<\/span>    (a) Already logged in via claude login? You're done.<\/span>\n#<\/span>    (b) Or generate a 1-year OAuth token for CI \/ non-interactive use:<\/span>\nclaude setup-token\necho<\/span> \"<\/span>CLAUDE_CODE_OAUTH_TOKEN=<paste>\"<\/span><\/span> ><\/span> .env\n\n#<\/span> 3. Verify<\/span>\naudit auth-check\n\n#<\/span> 4. Run<\/span>\naudit run --repo \/path\/to\/target --run-id my-run\naudit status --run-id my-run\naudit report --run-id my-run --format md ><\/span> report.md<\/pre>\n<\/div>\n
    By default the agent uses subscription billing<\/strong> via your Claude.ai
    \nlogin \u2014 it does not<\/strong> call the metered Anthropic API. The on-disk auth
    \nmodule scrubs ANTHROPIC_API_KEY<\/code> from the environment so it can’t
    \nsilently route around the OAuth flow.<\/p>\n
    \n
    Using a different model \/ provider<\/h2>\n
    <\/svg><\/a><\/div>\n
    The auth module picks one of three modes, in this order:<\/p>\n
      \n
    1. LLM gateway<\/strong> (OpenRouter, custom proxy, etc.) \u2014 when
      \nANTHROPIC_BASE_URL<\/code> points away from anthropic.com<\/code> AND
      \nANTHROPIC_AUTH_TOKEN<\/code> is set. The gateway env is left intact;
      \nonly ANTHROPIC_API_KEY<\/code> is scrubbed (it would otherwise outrank the
      \ngateway token).<\/li>\n
    2. Subscription OAuth (headless)<\/strong> \u2014 CLAUDE_CODE_OAUTH_TOKEN<\/code> from
      \nclaude setup-token<\/code>. Best for CI.<\/li>\n
    3. Subscription OAuth (interactive)<\/strong> \u2014 ~\/.claude\/.credentials.json<\/code>
      \nfrom claude login<\/code>. Best for local dev.<\/li>\n<\/ol>\n
      OpenRouter exposes Claude-compatible Anthropic-API endpoints behind its
      \nown credit system; that lets you spend OpenRouter credits instead of an
      \nAnthropic subscription, and gives you access to Sonnet\/Opus and<\/em> other
      \nmodels through the same SDK path. See       OpenRouter’s Agent SDK guide<\/a>.<\/p>\n
      \n
      export<\/span> ANTHROPIC_BASE_URL=\"<\/span>https:\/\/openrouter.ai\/api\"<\/span><\/span>\nexport<\/span> ANTHROPIC_AUTH_TOKEN=\"<\/span>$OPENROUTER_API_KEY<\/span>\"<\/span><\/span>\nexport<\/span> ANTHROPIC_API_KEY=\"<\/span>\"<\/span><\/span>           #<\/span> must be explicitly empty \/ unset<\/span>\n#<\/span> optional: pick a non-Anthropic model<\/span>\nexport<\/span> ANTHROPIC_MODEL=\"<\/span>anthropic\/claude-sonnet-4-6\"<\/span><\/span>\n#<\/span> or e.g.: ANTHROPIC_MODEL=\"openai\/gpt-5\"<\/span>\n#<\/span>         ANTHROPIC_MODEL=\"google\/gemini-2.5-pro\"<\/span>\n#<\/span>         ANTHROPIC_MODEL=\"qwen\/qwen3-coder-480b\"<\/span>\n\naudit auth-check                       #<\/span> confirms \"using LLM gateway at https:\/\/openrouter.ai\/api\"<\/span>\naudit run --repo \/path\/to\/target --run-id orun --max-cost-usd 30<\/pre>\n<\/div>\n
      Caveats:<\/p>\n