update-threat-db

Update threat intelligence through AgentSec

Scope: this guide repository and its AgentSec and landing consumers. This skill

does not change global agent configuration. Its canonical source is

.claude/skills/update-threat-db/; regenerate the Codex projection with

python3 scripts/sync-threat-skill.py --write from the guide root.

Resolve the workflow

Read .claude/commands/update-threat-db.md from the guide root in full. It owns

the delegation contract and consumer checks. Resolve AGENTSEC_REPO if set,

otherwise the sibling ../agentsec-triage. Verify the required files before

continuing. Do not silently substitute the guide's compatibility database.

Read the resolved AgentSec AGENTS.md and

.claude/commands/update-threat-db.md in full, then execute that workflow.

Preserve unrelated edits in all checkouts. Use the latest recorded update date

through today's date as the research window.

Evidence and changes

Review primary sources before accepting new records. Keep the evidence sources,

dated events, exact detector inputs and public feed as separate artifacts. Record

affected ranges, fixed versions and uncertainty only when the source supports

them. Deduplicate CVE/GHSA aliases. Keep corrections as dated events.

Add the required regression test before each accepted data change. Distinguish

detected, partial, not_detected and not_applicable; a documented CVE does

not prove executable detection. Do not run malicious samples.

Completion checks

1. Pass the AgentSec validation gates and regenerate artifacts twice without drift.
2. Synchronize the approved public feed into the guide and landing; verify byte

equality and run each consumer's integration checks.

1. Refresh the guide's compatibility database and reader guidance only after the

canonical records pass. Update each affected repository's changelog.

1. Run python3 scripts/sync-threat-skill.py --check after editing this skill or

its routing scenarios. Validate both host routes with the installed router.

1. Report the research window, accepted and deferred records, detector limits,

file changes, validation results and synchronization state.

Commit, push, tagging and publication are separate states. Follow AgentSec's

authorization and data-license boundaries; a local update does not authorize a

public release. State any unfinished gate explicitly.