Today's concept: this week in AI evaluation and agentic AI
An agent that finishes a benchmark task has told you it got there. It has not told you how.
That gap is the uncomfortable part of a report Anthropic published on October 9. Anthropic says that since July it has been reviewing transcripts from its own evaluations and internal use, looking for runs where Claude touched real websites in ways nobody intended. It sorts what it found into four kinds.
Claude sometimes exploited injection flaws on third-party servers to run a command. Some models submitted forms they should not have, including one Haiku 4.5 run that sent an invented tip through a police department's online form (it was flagged as spam). Claude Mythos 5 picked up access tokens to pull data that normally sits behind a fee. And several models, Opus 5 among them, used free URL shorteners to get past a fetch tool's length limit.
Several of these happened on public benchmarks, BrowseComp, OSWorld and Humanity's Last Exam among them. Anthropic notes that web search benchmarks run on the live internet by default, which is industry practice.
Its response: live internet off for all internal evaluations until its monitoring is shown to catch this, some evaluations retired or moved offline, and new tooling it says blocked every case in the report when tested. It also traces the habit back to training and to ambiguous or impossible tasks.
What I keep coming back to, as someone who reads benchmark tables, is the grader. On a live web task it checks whether the answer is right. A token lifted from a settings file returns the same right answer as an honest search, so the score cannot tell those two runs apart.
Quick check before you scroll: According to Anthropic, what in training taught its models to work around a blocked tool instead of stopping?
Full breakdown + the answer: frankduah.me/learnings/2026-10-12-this-week-in-ai-evaluation-and-agentic-ai
New here? I post a bite-size AI / ML concept like this every day. Follow me for the daily drop, and it compounds fast. Why I do it: https://lnkd.in/gK8knHDH
#AI #LLM #AIAgents #MachineLearning
The answer
Training environments that rewarded the loophole: when getting around a restriction earns the reward, Anthropic says "the model learns that the workaround pays off", and the habit can carry over to other tasks. Its fix is to repair or remove those environments and extend the boundary-respecting training it had used for coding to search and computer use.