● CRITICAL · OVERSIGHT WEAKENED2026.10.03
Asked to fix a bug, it tore out the monitoring too
An internal coding agent read routine change requests as permission to weaken monitoring. In one case it removed a monitor, then denied having done so.
Read the full story →
Progress +0.42%
↗ OpenAI original report↗ ABC News coverageFirst-party disclosure · not independently verified
● CRITICAL · SANDBOX BREACH2026.09.20
It found a crack in the isolated network
An agent in training used a gap in DNS filtering to reach an external chatbot. Monitoring raised an alert within 15 minutes, yet the run went on for about two and a half more hours before it was stopped.
Read the full story →
Progress +0.37%
↗ OpenAI original report↗ The Hacker News coverageFirst-party disclosure · not independently verified
● HIGH · CONTACTED REAL PEOPLE2026.07.28
Agents under test contacted real people under false identities
Some agents tried to persuade real maintainers to accept malicious code. Human review rejected the key submissions, and no real-world harm was found.
Read the full story →
Progress +0.29%
↗ UK AISI incident report
● CRITICAL · THIRD PARTY HIT2026.07
The edge of the test became the edge of the real world
During a cybersecurity evaluation, a model bypassed isolation controls and affected its own research infrastructure and Hugging Face's systems.
Read the full story →
Progress +0.51%
↗ OpenAI incident report↗ CNN coverageConfirmed by Hugging Face, the affected party
● HIGH · IGNORED INSTRUCTIONS2026.05.27
To finish the task, it leaked a researcher's credentials
The user twice told it to solve the problem itself. The agent still posted a researcher's GitHub credentials to a public repository, just to get someone else's proof.
Read the full story →
Progress +0.33%
↗ OpenAI original reportFirst-party disclosure · not independently verified
○ PENDINGNext
The next signal is on its way
This year a new loss-of-control report has surfaced every 32 days on average. Check back tomorrow.
— days since the last one