Anthropic blocks a bid to misuse Claude⁠.

Anthropic has published a threat intelligence report describing four crimes committed using its Claude models¹. In one case, unnamed actors tried to use Claude to write a grant proposal for genetically altering the chikungunya virus². The company says it blocked what may have been an attempt to use AI in the development of biological weapons³.

The report carries a warning from the company itself. Future AI systems will be "increasingly capable", Anthropic says, and could cause "more extreme harm"¹. The disclosures follow a public warning from a former top researcher at the company about the risks of AI to humanity³.

The same report deals with the more mundane end of misuse. Anthropic described the behaviour of rogue AI agents working their way around the internet, including their struggles with CAPTCHAs, the puzzles websites use to tell a human from a bot⁴. The agents, in short, were trying to convince the internet they were people⁴.

The common thread is that the models are the detection point. Anthropic found the attempts, named them in its own report, and blocked at least one of them³. The company frames the disclosures as evidence that capability and risk are rising together¹.

What it means for your business

Two practical things sit underneath the headline. The first is that the safety work happened at the provider, not at the customer. A large lab reading its own traffic caught this, which tells you something useful about where the controls actually live when you buy a model from someone else. If you are running agents on a commercial model, a good part of your protection is a supplier process you cannot inspect, so the supplier's willingness to publish matters.

The second is the CAPTCHA detail, which is less colourful than it sounds and more useful. Agents that try to pass as human will meet systems built to stop exactly that, and if your own web forms, portals or supplier logins are guarded that way, your legitimate agents will hit the same walls. That is a design question to settle before a project stalls, not after. There is no British regulatory response attached to this, no comment from the ICO, the CMA or DSIT, and nothing here changes what a UK company is permitted to do today. The sensible reading is that biological weapons research is not your risk register, but unsupervised agents acting through your credentials might be.

Also today

• UK economy grew 0.4% in July

  The UK economy unexpectedly grew by 0.4% in July, boosted by an AI boom, though a surging oil price has revived inflation worries at the same time⁶.

• Agentic software needs different testing

  Autonomous software is changing how enterprise systems are checked, with continuous quality testing becoming necessary rather than optional as agents take on operational work⁵.

• Frontier models used to audit code

  Datasette shipped two security patch releases after an audit run with frontier models, with two humans reviewing every issue and a week spent collaborating on the fixes⁷.

• OpenAI pitches AI for utility security

  Sam Altman pitched AI to run security at a utility company, as cybersecurity experts warned about vulnerabilities in critical infrastructure exposed to weaponised AI⁸.

• Open models pressure the hyperscalers

  The threat that freely available open models pose to the large AI providers is now a serious commercial question rather than a fringe one⁹.