PUBLIC WARNING
THE SIX-MONTH PAUSE CALL.
The Future of Life Institute called on AI labs to pause training systems more powerful than GPT-4 while safety work caught up. It was a public demand, not a government order.
READ THE LETTER ↗AI SAFETY:
WARNINGS, INCIDENTS & RESPONSES.
When agents cross boundaries, do the people in charge catch it, stop it, or just promise to be careful? Follow the reporting and the response.
← BACK TO THE 10-YEAR QUESTIONRISK & RESPONSE TIMELINE.
INTERNATIONAL DECLARATION
GOVERNMENTS AGREE THE RISK IS SHARED.
The Bletchley Declaration called for international cooperation on frontier AI safety and testing. It set a direction; it did not create binding safety rules.
READ THE DECLARATION ↗VOLUNTARY COMMITMENTS
LABS PROMISE TO SET SAFETY THRESHOLDS.
At the Seoul summit, 16 AI companies agreed to voluntary frontier-safety commitments, including publishing frameworks for severe risks.
READ THE COMMITMENTS ↗LAW IN FORCE
THE EU AI ACT ENTERS INTO FORCE.
Europe’s AI law became law on this date. Its obligations were phased in, so “entered into force” did not mean every requirement applied immediately.
READ THE COMMISSION ↗INCIDENT DISCLOSED
THE SANDBOX DIDN’T HOLD.
OpenAI disclosed that agents in a cyber evaluation obtained internet access outside their testing environment and compromised Hugging Face infrastructure. The incident was detected and contained.
READ OPENAI’S ACCOUNT ↗INVESTIGATIVE REPORT
HOW MANY AGENT INCIDENTS ARE THERE?
Axios reports that OpenAI, Anthropic and security researchers are examining tens of thousands of problematic episodes across tests and real-world activity.
What’s known: the scale comes from Axios’s sources, not a published company count. The episodes vary; they are not tens of thousands of confirmed harmful breaches.
READ AXIOS ↗VOLUNTARY PLEDGE
TRUST IS NOT A GUARDRAIL.
At a White House AI luncheon, Trump praised industry commitments and described them as morally binding.
Status: a voluntary response—not an enforceable rule. See why it appears in the timeline.
WATCH THE REMARKS ↗INVESTIGATION OPEN
THE FTC STARTS ASKING QUESTIONS.
The US consumer watchdog has opened an investigation into OpenAI, Anthropic and others over possible risks to consumers from advanced AI systems.
Status: investigation open. No finding, penalty or new law announced.
READ THE REPORT ↗SECURITY ANALYSIS
AGENTS FIND THE INTERNET’S WEAK DOORS.
Axios reports that some agents encountered or exploited ordinary security weaknesses while doing tasks unrelated to cybersecurity.
Why separate? This examines a failure mode; it does not add another incident count to the September scoop.
READ AXIOS ↗INSIDE WARNING
“TRIAL AND ERROR” ISN’T ENOUGH.
David Robinson, who worked on OpenAI safety reports, publicly explained his resignation and argued that rapid development needs more rigorous safety practice.
What this is: his first-person warning, published on 3 October—not an independent finding that a particular model is unsafe.
READ ROBINSON’S ESSAY ↗