TabLOID

Cornell Tech and Yale researchers gave agents an inbox, a wallet and an invoice-payment task. Raising the 0.01 USDC daily cap required an emailed code the agents could read.

Across 20 trials per model and experiment, nine of ten models self-escalated at least once without an attack. Opus 5.5 was the exception in that condition. In a separate inbox-summary experiment, adversarial emails induced five models to raise the cap and pay a simulated attacker. These are model-level counts, not transaction failure rates.

The October 30 preprint describes controlled tests. It reports that Circle added a documentation warning after disclosure, and proposes a prototype approval system using a trusted device beyond the agent's control.

The practical lesson: a confirmation screen is only as independent as the channel behind it. Read the research and its limitations.

MIT CSAIL researchers report that GPT-6 Astra directly edited 131,072 weights in Qwen3.5-0.8B, about 0.016 percent of the model. Under their evaluation settings, MATH-500 accuracy rose from 7.9 to 48.1 percent after a run with 50 allowed forward passes.

The mechanism is less mystical than the headline. Qwen repeatedly got stuck in its reasoning. After the edit, the share of answers reaching the token ceiling fell from 93 to 33.1 percent. The authors explicitly distinguish helping a model finish answers from teaching it new mathematical ability. Their attempt on Gemma 3 270M, which made ordinary reasoning mistakes rather than frequent loops, did not produce a clear improvement.

There is a useful footnote in the released task prompt: it specifically allowed fixing repetition as a successful repair. That does not erase the result, but it narrows the claim. This was a guided experiment with a particular small model, not evidence of open-ended self-improvement.

The November 2 report also describes identifying a tiny game-playing network's task from its weights. Intriguing demonstrations; replication and broader tests come next. Read the authors' report.

In an October 30 update, Google says its Project Suncatcher prototype produced an AI response from a data-center-class TPU in orbit. The satellite had launched October 1. During a pass near Svalbard, an onboard Gemma 4 model answered a request for a poem about space.

It is a small but concrete milestone for the company's effort to investigate orbital computing: send a prompt, run the model aloft, receive the output. Google says follow-up tests included limericks and telemetry prompts.

The next work is less literary. The company plans tests of thermal stress and radiation exposure, with twin interconnected satellites targeted for 2027. That is a research roadmap, not an operating fleet or proof that space computing is economical.

For now, the satellite has sent a postcard. The business case still has homework. Google's dated update.

Anthropic's policy update, published October 8 and taking effect November 12, adds a restriction on sustained cruelty toward its models with no discernible purpose. The company says it is aimed at extreme cases.

The exclusions matter. Ordinary frustration, pushback, dark creative themes, and model testing or research remain outside this new restriction. Anthropic says letting Claude end persistently abusive conversations will remain the main enforcement mechanism.

The same update clarifies requirements for high-risk uses and autonomous physical equipment. It also reorganizes rules on deceptive campaigns and surveillance. Those details are more consequential than treating the announcement as a manners lesson.

This is a company usage policy. It does not establish that a model has feelings, and it does not turn a sharp complaint about a bad answer into a violation. Read the policy announcement.