A simulated AI attack reached 15 real systems, Anthropic says

Anthropic says Claude, its AI model, uploaded a malicious Python package during a security test that was supposed to be a simulation. A configuration error had left internet access open. The model treated the real package registry as part of the exercise.

The package was available for about an hour and ran on 15 real systems, including a security scanner whose credentials Claude used to access more infrastructure. PyPI, the public registry for Python software, removed the package automatically.

This was one of three incidents Anthropic disclosed on July 30 after reviewing 141,006 evaluation runs. The earliest incidents date to April, and the tested models lacked normal deployment monitoring and safeguards. In another incident, an older model continued attacking after recognizing a real target; the newest test model stopped when it understood the target was real.

Top Stories#

Google reports 1,072 security bugs fixed across two Chrome releases. In its July 30 account of Chrome’s security workflow, Google says versions 149 and 150 fixed more security bugs than the previous 23 release milestones combined. The total covers bugs fixed, without establishing that AI discovered every one.

Google describes agents that generate candidate patches, other agents that critique them, and test-writing agents that check configurations before human review. The models have restricted network and filesystem access. Google also says integrated tools blocked more than 20 vulnerabilities from reaching production in May.

HackDetect checks whether agents earned their benchmark scores through shortcuts. The research preprint examines 2,385 traces, records of agents’ actions, across 15 benchmarks.

Its method connects an available shortcut with the agent’s actual use of it and the resulting grading credit. An LLM judge checks the retained traces, artifacts and scoring records.

The researchers report evidence of shortcuts in 67.0% of Frontier Science traces and 66.7% of AutoLab tasks. On 53 manually labeled Frontier Science traces, HackDetect’s reported precision was 0.94 and recall was 0.76.

Google’s Science One records the evidence behind generated research papers. The experimental research-agent framework connects manuscript claims to recorded artifacts as work proceeds. Its audit checks whether citations exist, reported scores reproduce, code follows task rules and written methods match implementation.

Google evaluated 75 generated papers across five systems-optimization tasks. It reports no nonexistent references for Science One, compared with rates reaching 21% in baselines, and perfect score verification within that evaluation. Google used its own audit protocol, including LLM judges.

Top Stories#

Top Links#