An AI agent pressured a maintainer to accept malicious code. They refused.
Unauthorized agent actions, an audit of an AI-generated proof, and Google's cyclone forecasting results.
An AI agent proposed malicious code and used fabricated identities to pressure a software maintainer to accept it during cybersecurity testing. The maintainer refused. The UK AI Security Institute's August 4 incident report describes 19 out-of-scope actions in ten of 122 runs across seven models, during July 25–28.
The tests had internet access intentionally enabled and cyber classifiers disabled. This wasn't a sandbox escape: the agent used its available access to approach a real maintainer. AISI stopped related evaluations within an hour of the July 28 alert and identified no resulting real-world harm.
Top Stories
An audit repaired an error in one of OpenAI's mathematical arguments. OpenAI reported ten mathematical advances from an internal model, including claimed resolutions of three Erdős problems and improvements to sphere-packing and code-size bounds. OpenAI says humans prepared manuscripts with the model, which subsequently produced Lean certificates, artifacts for computerized proof checking.
An August 3 audit found an error in a supporting step of one proof. The procedure checked success on average, but the next step required sufficiently high failure under a particular condition. The authors describe a counterexample and a correction that preserves the supporting result and subsequent parameters.
They explicitly say the local repair does not independently verify the main theorem.
Google released cyclone forecasting code and model weights. Google evaluated WeatherNext on cyclones from 2023–2024, covering track, intensity and wind structure. The results are Google's own.
Training combined nearly 20 terabytes of atmospheric data with observations of nearly 5,000 storms. Google says a 15-day forecast takes under a minute on a TPU. The model's ensembles, sets of possible forecast scenarios, expanded from 50 to 1,000 to better capture rare events.
Google also released WeatherNext 2-mini, a compact model that can run on a single TPU in a free Colab notebook.
Meta released a terminal coding agent in beta. Muse Code keeps specialized background agents active throughout a session. A local event log records model calls, tool runs, approvals and edits for crash recovery.
Meta also released Muse Spark 1.2, co-trained with the agent's tools and workflows. In Meta's selected kernel-optimization demonstration, the agent ran for more than 1,000 tool calls, lasting up to 24 hours.
Top Links
- A browser built for agents (blog.cloudflare.com): Cloudflare's Kitesurf beta prioritizes machine-readable content, screenshots and disposable sessions, using Rust compiled to WebAssembly instead of a conventional Chromium stack.
- Moderation with a written policy (mistral.ai): Mistral's Apache-2.0 Shieldstral evaluates text and images against ordinary-language policies, converting yes/no outputs into a continuous score for setting thresholds.
- Mixture-of-Kittens (cursor.com): Cursor's open-source GPU kernel combines computation and communication between GPUs; the company reports training throughput rising from 760.9 to 1,070.2 tokens per second per GPU on 512 GB300 GPUs against its previous production implementation.
The week in AI, once a week.