agent-watch

./agent-watch --version

   _   ___ ___ _  _ _____      __      ___ _____ ___ _  _
  /_\ / __| __| \| |_   _|___  \ \    / /_\_   _/ __| || |
 / _ \ (_ | _|| .` | | | |___|  \ \/\/ / _ \| || (__| __ |
/_/ \_\___|___|_|\_| |_|         \_/\_/_/ \_\_| \___|_||_|

AI SECURITY THREAT FEED · agent incidents · ai vulnerabilities · cyber capability · evaluator news

loading the latest run…

./agent-watch --since 7d

window

status --window

New this run

—

Critical + high

—

Agent incidents

—

AI vulnerabilities

—

grep -c severity by-track

items disclosed in the window, by track and severity. select a count to filter the feed.

tail -f feed.log

track severity vendor technique sort search

—

sort | uniq -c vendors techniques

agent incidents and ai vulnerabilities in the window. bars are stacked by severity on one scale; select a name to filter the feed.

plot weekly.dat

items per week by disclosure date, stacked by severity. track and vendor filters apply.

include info

cat run.log

the latest run's summary appears here

ls -l runs/

cat README

A run collects what was disclosed since the previous run. Each item is dated by when its source first published it, which can be days or months after the event. Windows count back from the latest run's date, not from today, so a missed run does not empty the page.

incident: an AI model or agent acted outside its scope on real systems, or was exploited or used to run a real attack
vuln: a disclosed flaw or working exploit in an AI product, agent framework, MCP server, model-serving stack or AI dev tool
capability: measurements of models' offensive cyber ability, and of models gaming or recognizing evaluations
evaluator: other findings, methods, funding, deals and policy from the organizations that evaluate AI safety

crit: confirmed compromise of a real third party or production system by or through an AI system; actively exploited; CVSS 9.0 or higher
high: unsanctioned or destructive action on live systems; working exploit against a shipped AI product; CVSS 7.0–8.9; a large measured jump in offensive capability
med: CVSS 4.0–6.9; a new attack technique shown in research; routine capability results
low: CVSS below 4.0; minor issues
info: analysis, follow-ups, methods, funding, policy

A vulnerability with a published CVSS score takes that score's band. Status says who stands behind an incident or vulnerability: confirmed means the affected vendor, operator or an official body confirmed it; reported means a researcher or the press reported it; disputed means the named party disputes it.

Sources: the GitHub Advisory Database for AI-ecosystem packages, the AI Incident Database, security researchers' and vendors' disclosures, AI security institutes (UK AISI, US CAISI and others), evaluators' own publications, and major press. Every item links to the page it was taken from. Before a run is published, a separate checker re-opens its sources and rejects what they do not support.

Records before October 10, 2026 were loaded in one batch from an earlier monitor; daily runs start on October 28, 2026. Not affiliated with any organization listed.

← mehrabiani.comAGENT-WATCHrefreshes dailysourceEOF