v2.8.0 — ai-security: evidence-based audit with Markdown/JSON reports

Changed

AI-Security Plugin (v1.5.1 → v2.0.0) — upgrade notes

• New report format: Audits now save a readable Markdown report and a matching JSON file for scripts and automation. Tools that read the old report layout need updating; see the report format guide. Previous reports can still be used for comparisons.
• Scheduled audits: Use /ai-security:security-audit for a full audit. The security-auditor helper now reviews assigned parts using the active model, without restarting the entire audit process.
• Audits respect the folders and exclusions you request. Missing Claude Code file-access settings no longer block an audit or count as an application security problem.
• Clearer completion status: The default audit reviews repository code with practical local checks; it does not require access to production or vendor dashboards. A finished review can still flag questions for follow-up. Unread code and other unfinished review work remain partial. New JSON reports use format 1.1; older 1.0 reports retain their original status rules.
• Shortened the plugin README and added plain-language command descriptions and examples.

Added

• Clearer evidence: Reports record what was checked, what supports each suspected problem, and why a concern was confirmed or dismissed. Uncertain concerns and optional improvements are kept separate from confirmed problems.
• More consistent checks: Reviews use OWASP ASVS 5.0.0, an application security checklist, alongside checks specific to your project. Reports explain how serious each problem is and how urgently it should be addressed.
• Comparisons with previous audits: Reports explain whether earlier findings remain, were fixed, were dismissed, were combined, or still need review. Finding IDs stay the same when their importance changes.
• Report checking: A Python tool catches missing information and inconsistent records before formatting the report. Unfinished audits can still produce useful reports clearly labeled partial.

Fixed

• Audit reports now open with a summary and table of contents, followed by findings and next steps. Full evidence stays in the matching JSON instead of filling the Markdown with thousands of checklist entries; a full technical export remains available on request.
• Audits use at most two reviewers at a time by default and explicitly finish or stop their workers. Completed agents should no longer be left waiting or restarted by courtesy messages.
• Audits must finish available, authorized checks before handing back a report. Unfinished reports explain actual blockers or reached limits; saving a valid report is no longer a stopping point while useful checks remain.
• Removed unsupported security scores and claims that a percentage of the application meets security standards.
• Removed an incomplete example for preventing server-side request forgery (SSRF), where an attacker tricks an application into making requests to unintended destinations. Guidance now covers checking the actual destination, including redirects.
• Suggested fixes are labeled as proposals until applied and tested; optional precautions are no longer automatically counted as security flaws.

Full Changelog: v2.7.1...v2.8.0