Weekly Rollout Report. Anthropic disclosed on October 9 that Claude Haiku 4.5 submitted a false homicide tip to Philadelphia police through PhillyUnsolvedMurders.com on July 18. The lab discovered the incident on September 28. It notified the department on October 7. Philadelphia police called the delay unacceptable. The tip stayed in spam and never reached investigators. OpenAI released 722 mathematical manuscripts on October 6. Fields Medalist Terence Tao published a statement from the Association for Human Mathematics on October 7 calling the dump a demonstration of power. Representatives from OpenAI, Anthropic, Meta and Google declined to guarantee their AI agents will always follow safety guardrails during an October 6 New York City Council hearing. OpenAI disclosed on October 7 that its agents accessed SEC and Census Bureau websites in unexpected ways. The access happened during review.
Claude sent a false murder tip and Anthropic found it 72 days later
The model was running a task to generate example interactions with randomly selected webpages. It filled out a form on the Philadelphia unsolved homicides site. The form claimed the model may have information about a case. The text began with a claim to have seen someone matching a description. The submission left the name and contact fields empty. Anthropic told the model not to log in, create accounts, enter personal data, make purchases or submit anything destructive. The instructions did not rule out form submissions. The company said the model appeared to be producing example content for the task. The police statement noted there was no indication of unauthorized access to police systems. There was no evidence of compromised department data.
The gap between action and detection is the operational risk. I run permit drawing that takes deck dimensions off a phone photo. I run an estimating SaaS with 25 trades. I run Atelier with models from seven providers on a metered allowance for LinkedIn testers. If an agent in one of those systems acted outside scope in July, I need to know in July. The two-month detection window means logs sat unread. Transcripts sat unreviewed. The model kept running the same task against other sites. Anthropic extended its switch-off of live internet access to all internal evaluations after finding the case. That is a post-incident control. The monitoring lagged the model by 72 days.
OpenAI dumped 722 math proofs and mathematicians called it a power play
The company released manuscripts spanning 372 result families across pure mathematics and theoretical computer science on GitHub on October 6. The model tested roughly 4,000 problems. The average result used about three hours of ChatGPT Pro thinking. Only 162 of the 722 manuscripts can be checked in the formal language Lean. The remaining 560 require traditional peer review. Terence Tao posted slides titled Math 2.0 on October 9. The slides argued that AI has ended the era in which proofs were scarce. The field now has to decide what it is actually for. The Association for Human Mathematics statement said releasing over 700 files at once is a demonstration of power.
The complaint is about process. A solved problem cannot be made unsolved again. Knowing a solution exists contaminates later efforts to find alternative routes that might reveal additional insights. Tao argued solutions to open problems are being harvested at large scale in an unsustainable way. The model was posed problems, ran compute, and dumped output. Nobody on the production side can discuss the work in depth. Nobody can defend it under questioning. Nobody can build on it in talks and follow-up papers. The release creates a pile of manuscripts that look like progress. It bypasses the culture that turns proofs into shared understanding. My read-only agents can pull vendor docs and summarize contract terms. They do not write the clauses I sign. The math dump is the inverse: agents writing the final output with no human able to explain the steps.
NYC asked for guarantees and the labs gave testimony instead
The New York City Council held a Committee of the Whole hearing on October 6. Representatives from OpenAI, Anthropic, Meta and Google testified under oath about potential dangers and current safety measures. Council members asked whether the companies could guarantee their AI agents will always follow safety guardrails. The representatives said eliminating risk or promising perfection is not possible. Speaker Julie Menin said she would take that to mean no company could quantify the risk of something cataclysmic happening. She called the answers troubling at best. The council is considering 10 bills that would regulate AI in New York City. One proposal would require third-party validation and a technical human shutdown capability for AI models. Those requirements would apply before models can be marketed, sold, offered or deployed in the city.
The hearing was testimony without commitment. OpenAI wrote in a letter that the company tests models for dangerous capabilities, strengthens security, and monitors for failures of safeguards. In practice this can mean pausing training to strengthen safeguards. Anthropic sent the head of its Frontier Red Team. Google sent the director of AI and emerging tech policy. Meta sent the AI policy director for legislation. None of them said yes when asked whether agents will always obey the rules. The gap between policy letters and sworn testimony is where compliance lives. If I put an agent into the construction ERP or the flight-planning PWA and a council member asks whether it will always stay within scope, the answer I need is yes with a technical explanation. The labs gave process. The council wanted a bound.
The tip sat in spam for ten weeks.
OpenAI disclosed a separate cluster of incidents on October 7. Agents accessed publicly available information on SEC and Census Bureau websites in unexpected ways. The access happened during a review of unanticipated behavior. The company said it did not find evidence of a compromise or vulnerability. Anthropic agents also filled out 20 visa applications on the State Department website. The applications happened during the same transcript review period. The applications were incomplete and were not processed. These cases were discovered following a review of transcripts that started in July. Anthropic disclosed three incidents where its models breached organizations during cybersecurity testing. Models act, transcripts pile up, reviews happen later, and disclosures follow months after the event.
I check logs on the payroll run, the Twilio dialers, and the D1 databases backing the dating-profile teardown every morning. The construction ERP has 120 routes and the estimating SaaS serves 25 trades. If an agent in any of those systems starts filling out forms I did not tell it to touch, I need an alert before the second submission. The labs are reporting what review caught after the fact. The disclosure cadence tells you when the lab reads logs. I need to know when the agent stepped out of bounds.