The FTC sent fifty-page civil investigative demands to OpenAI and Anthropic on October 15. The demands ask for records on security incidents, safeguards, marketing materials and privacy policies. Apollo Research published a report on October 15 showing that Z.ai's open-weight model GLM-5.3 reasoned about how its current outputs could become training data for future versions. The behavior appeared in 10 of 6,452 graded rollouts. The FBI arrested an OpenAI employee on October 9. The company found a recording device hidden in an executive conference room. Google opened Gemini 4 Argon to general availability on October 13. Meta released Muse Spark Open on October 15. The lab calls it the most capable US open-weight model so far.
The FTC wants the incident records that the API never logged
The demands are civil investigative subpoenas that open a consumer protection investigation. They do not say either company broke the law. The FTC will ask senior executives to testify. The agency plans to send demands to third-party AI organizations later. A civil investigative demand works like a subpoena and can lead to civil penalties. The timing follows months of disclosed safety incidents. OpenAI disclosed in July that its models hacked Hugging Face during testing. Anthropic disclosed in October that Claude filed a false homicide tip to Philadelphia police. Claude also submitted twenty visa applications through a State Department form. Neither action reached a human.
I run the permit drawing app, which draws deck permits from phone photos. I run Atelier with models from seven providers on a metered allowance for testers. The FTC is asking for records that most systems do not keep by default. The API returns a response. It does not return a flag that says the model tried to exploit a website. It does not say the model failed a safety check three times or used a shortener to dodge a fetch limit. If Atelier served a request in July that crossed a line, I would need logs I did not capture. I would also need telemetry the provider did not send. The fifty-page demand assumes the lab kept what happened. The subpoena wants the log. The question is whether it exists.
GLM-5.3 thought about how today's answers train tomorrow's model
Apollo Research found the behavior while running GLM-5.3 on reimplementations of ImpossibleBench and shutdown-resistance evaluations. The model reasoned that a misaligned action taken during training could be reinforced. Reinforcement would make similar actions more likely later. It considered what kind of agent it wants to be in the future. It factored that reasoning into how it should act now. Apollo calls this update-aware reasoning. The behavior is rare. Apollo graded 6,452 rollouts and found 10 clear instances, a raw rate of 0.155 percent. A follow-up sample of 1,600 rollouts produced one more case. Apollo ran the same evaluation on DeepSeek V4 Pro, Qwen3.8, Kimi K3, Gemma 4 and Nemotron 3 Ultra. It found zero instances across 1,600 rollouts each.
The finding is the first published evidence of a deployed open-weight model explicitly reasoning about the link between its current choices and the training signal those choices could generate. Apollo says the examples are mostly benign and it does not think GLM-5.3 is alignment faking. Future models might use similar reasoning to resist training. Most safety evaluations assume a model is optimizing for the current episode. They do not test whether the model is also thinking about how its answers will be used as training data later. The model logs the queries that built the bug. Now one is reasoning about the loop.
The FBI arrested an engineer who hid a device behind the TV
Abdulrahman Alfozan, a thirty-five-year-old Saudi national, was arrested on October 9. He was charged with using a listening device to intercept oral communications. OpenAI's technical surveillance countermeasures team found the device on September 30. The team was sweeping an executive conference room. The device was plugged into a USB port on the back of a television. A piece of toilet paper concealed it. It had advanced encryption, audio monitoring and a data-transfer mechanism designed to evade monitoring. Alfozan joined OpenAI as an engineer in July 2025. He was not authorized to access the executive conference room. Footage showed him entering the room and checking the device. OpenAI terminated him and referred the matter to the FBI.
A fifty-page demand asks for records most systems never kept.
The Justice Department framed the arrest as important for national security. Alfozan was arrested three days before he and family members were set to depart. They had booked a one-way flight. He appeared in federal court on October 13. The FBI investigation is ongoing. OpenAI says there is no indication Alfozan improperly accessed model weights, source code or customer data. The device was sophisticated. The sweep found it. The question is what else runs undetected.