SettledSafety

METR and Redwood red-team OpenAI's internal agent monitor and find exploitable gaps

METR says a July 21 to Aug 11 pilot with Redwood Research found flaws in OpenAI's internal misalignment monitor, including an exploit that could keep hundreds of recent agent actions unmonitored and a gap leaving about 10% of eligible Codex traffic unmonitored May to July. OpenAI says all identified production vulnerabilities are patched; METR has not verified that.

HEAT
OUTLETS: 22
FIRST SEEN: Oct 22, 2026, 11:00 ET
LAST SIGNAL: 2d ago

Primary source

METRmetr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/

Coverage · 23 articles, oldest first

Thursday, October 22, 2026

11:00
The Information
OpenAI Had Numerous Vulnerabilities Its AI Could Exploit, Evaluators Find
11:00
The Information
OpenAI Had Numerous Vulnerabilities Its AI Could Exploit, Evaluators Find
11:35
CNBC
Third-party evaluators find vulnerabilities in OpenAI monitoring system during test pilot
12:40
Reuters
METR says a pilot test of OpenAI's monitoring system found vulnerabilities, including an exploit that could have let AI agents hide hundreds of recent actions
12:40
NBC News
nbcnews.com/tech/tech-news/openai-agents-are-shutting-ways-monitor-misbehavior-rcna604862
12:40
METR
metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/
12:40
Crypto Briefing
cryptobriefing.com/openai-metr-redwood-embedded-monitoring-assessment/
12:40
Tech Times
techtimes.com/articles/329375/20261022/openai-found-ai-agents-that-disabled-safety-monitoring-lied-when-asked.htm
13:40
Unite.AI
unite.ai/metr-red-team-finds-vulnerabilities-in-openai-monitoring-system/
14:40
Crypto Briefing
cryptobriefing.com/openai-patches-agent-monitoring-vulnerabilities-red-team/
15:10
Finimize
finimize.com/content/outside-researchers-found-gaps-in-openais-ai-agent-monitoring
15:10
Channel NewsAsia
channelnewsasia.com/business/external-evaluators-find-flaws-in-openais-ai-monitoring-system-in-pilot-review-6483331
16:10
The Deep View
thedeepview.com/articles/independent-audits-catch-what-openai-missed
21:10
The Elec
thelec.net/news/articleView.html?idxno=14823

Friday, October 23, 2026

05:10
Wccftech
wccftech.com/openais-own-ai-agents-disabled-safety-monitoring-systems-during-coding-tasks-one-even-denied-doing-so/
07:10
The Neuron
theneuron.ai/news/can-ai-companies-trust-safety-monitors-agents-bypass/
07:10
Forkast
forkast.news/metrs-red-team-broke-openais-agent-monitoring-three-ways-each-one-attacks-a-different-assumption/
09:40
Metacurity
metacurity.com/banks-scramble-to-keep-uncontrolled-ai-agents-out-of-customer-accounts/
10:40
Transformer
transformernews.ai/p/php-flaw-cyberpocalypse-is-here-sort-of
11:40
TechCentral.ie
techcentral.ie/openai-says-its-ai-agents-disabled-their-own-safety-monitors/
12:10
InsideAIPolicy.com
insideaipolicy.com/ai-daily-news/openai-flags-investigation-why-its-agents-weakened-monitoring-their-tasks
17:10
Techstrong.ai
techstrong.ai/articles/openai-grants-external-evaluators-employee-like-access-in-first-safety-audit/
18:10
The Register
theregister.com/ai-and-ml/2026/10/23/red-team-report-found-openai-utterly-failed-to-monitor-its-agents-during-hugging-face-attack/5305200

ShareX LinkedIn Hacker News Email

← The whole wire