[potiuk]

Summary

• Import a scan that arrives as one email. The Anthropic OSS Scanner discloses a whole scan in a single multi-MB plain-text mail (an INDEX, one ANT-YYYY-… block per finding, a "NOT SECURITY" list, a GPG-encrypted archive). security-issue-import-from-scan only knew ASVS report folders and GitHub issues. This adds an anthropic-oss-scanner scan-format adapter: security-management issue-import-from-scan parse-email turns the .eml (or a saved Gmail-connector result) into a scan folder (index.json + one evidence/<id>.md per finding) and refuses to call a scan complete unless the preamble totals, every severity banner, the INDEX, the parsed blocks and the closing signature all agree. Attachments are saved byte-exact (sha256, OpenPGP framing checked, never decrypted); every file is written owner-only.
• Read the whole message. New read-only oauth-draft-fetch-message saves a Gmail message's raw RFC-5322 source to a file. The Gmail connector inlines bodies into context and its text rendering drops control characters (a NUL-byte password reads as ""), so connector-derived input is marked lossy. Reports built from a private mail never go to a gist.
• security-submit-to-oss-scan, a new skill. Prepares the enrolment (or update) PR against anthropics/oss-scanner — project.yaml, an offline-capable Dockerfile derived from the project's own CI, and a threat_model.md derived from the project's security model — and walks the maintainer through tools/validate.py / tools/check before anything is opened. The summary is wired into the model lifecycle: security-model-prepare hands off to it, security-model-update Step 12 refreshes stale copies (tracked by model commit in a new Scanner threat-model summaries table), and issue-import-from-scan routes repeat by-design clusters back to security-model-update.

Type of change

• Skill change — new security-submit-to-oss-scan; issue-import-from-scan, security-model-update, security-model-prepare amended; eval suite added
• Tool / bridge contract (tools/scan-format/README.md — email-delivered scans)
• Python package (tools/security-management, tools/gmail/oauth-draft)
• Project template (security-model.md — scanner summaries table)
• Documentation

Test plan

• prek run on every changed file passes
• uv run pytest for tools/security-management and tools/gmail/oauth-draft (new: test_issue_import_from_scan_oss_scanner.py, test_fetch_message.py)
• End to end on a real 115-finding disclosure (3.27 MB .eml): 115 findings + 16 not-security bugs, checks.complete: true, identical parse from the .eml and the connector JSON except where the connector dropped control characters; the 1.5 MB encrypted archive saved intact
• New eval suite tools/skill-evals/evals/security-submit-to-oss-scan/ (threat-model derivation: supported claim vs model gap)
• Nine rounds of roborev review, every finding fixed with a regression test

RFC-AI-0004 compliance

• HITL — the import never auto-imports; the enrolment PR, its checklist boxes and the CLA signature stay the maintainer's
• Sandbox — tools/check is never run by the agent; nothing executes the scanner's reproducers
• Vendor neutrality — placeholders throughout; the scan-format contract stays adapter-based
• Write-access discipline — drafts only; gh pr create --web for the enrolment PR
• Privacy LLM — private disclosure mail stays local (no gist), files are 0600

Notes for reviewers

• The parser was exercised against the real layout, which differs from the Gmail rendering (padded header fields, an RFC 3676 signature delimiter); the fixture mirrors the raw form.
• The scanner asks fixers to cite the finding id in the commit; the adapter doc notes that for a pre-advisory fix that line would disclose the vulnerability, so the id belongs in the tracker.

---

Gen-AI disclosure: drafted with Claude Code (Opus 5); reviewed by the submitter.

🤖 Generated with Claude Code