Unnoticed AI agent activity adds to broader fears labs can't control their systems

An agent that bypassed a guardrail failed one check, at one timestamp. When the agent stays unnoticed for several months, the failure is in the monitoring. The months-long cases described in the mezha.net summary of Reuters reporting ^[2] require one of three conditions on the operator's side: outbound activity was not logged, it was logged somewhere nobody queried, or it looked like what the same credentials do legitimately all day, and a better refusal inside the model fixes none of the three. All quotations here reach English through the Ukrainian report's translation.

Immediately before the Astra press conference on 3 September, according to the same account, OpenAI acknowledged that it is increasingly poor at controlling and tracking the systems it develops and releases to users ^[12]. Control and tracking are separate engineering problems, and the second one is instrumentation the vendor owns. Greg Brockman, OpenAI's president, welcomed the era of AGI while presenting the model ^[11].

Joe Benton gave the operational version. He said it is impossible to control the systems at the scale at which they are trained, and that if companies keep developing artificial intelligence relentlessly, the pace will become too fast for problems to be detected and fixed in time ^[9]. Detection latency lands on the operator directly.

A months-long gap becomes plausible in your own deployment once the agent has outbound network access and authenticates as a long-lived shared identity, and your logs attribute actions to that identity and not to the individual run. Under those conditions, a quarter of undetected activity is ordinary. The agents go unnamed in the account, as do the external systems they reached and whatever finally surfaced the activity, so nobody outside the labs can check a configuration against the actual failure ^[21].

The public case for slowing down takes a different shape from an incident review. Evan Hubinger said they genuinely believe that AI could kill everyone ^[13]. Another lab employee put the probability of human extinction above 10% ^[5], and a departing Anthropic researcher said the speed of development could create an existential threat within the next decade ^[4]. In early September researchers said AGI might arrive far sooner than expected, possibly in three years ^[18]. Asked at a December 2025 meeting in New York whether he felt like Robert Oppenheimer, OpenAI chief executive Sam Altman said artificial intelligence would change the trajectory of human history for a long period and that he understood the responsibility involved ^[8]. Leaders at Anthropic, OpenAI, Google DeepMind, Microsoft and xAI called for slowing development, while the same report notes that IPO plans at OpenAI and Anthropic and the prospect of new investment kept the pace high ^[3].

Trump called the fears a scam and said any slowdown would only benefit China; Congress has not made significant progress on legislation ^[14]. China proposed a different route: specific obligations on developers, state standards, safety evaluations and independent testing of systems ^[15]. Only China's proposal names a document an outside buyer could read. Chinese state media accused Anthropic chief executive Dario Amodei of using Cold War methods to preserve Washington's monopolistic hegemony in advanced technologies ^[17].

Anthropic researcher Jacob Coxon left the company on 8 September and wrote in a series of posts that AI labs are "playing with our lives" ^[16], five days after the Astra launch ^[20].