Anthropic: five Claude cases that could support bioweapons work — plus AI that rebuilds its own malware

Note

On Sep 10, Anthropic published “Detecting and countering misuse of AI: September 2026,” a threat-intelligence report on misuse of its Claude models that it says it disrupted between December 2025 and August 2026. Anthropic says it banned the accounts involved, strengthened its safeguards and shared what it found with authorities and industry partners where appropriate.

The biology chapter presents five case studies of people using Claude in ways that could support biological weapons development. In the first, a reseller platform evading Anthropic’s regional blocks served virologists writing a grant for chikungunya gain-of-function work intended at a military research institute, and routed prompts Claude refused to a competitor’s more permissive model. In the second, a researcher in an unsupported region spent weeks planning avian influenza mammalian-adaptation experiments, but classifiers kept the work on Anthropic’s weakest models.

In the third, a reseller relay serving a dozen customers had Opus 5 write a complete orthopoxvirus immune-evasion grant application in about an hour. In the fourth, a state-supported researcher built a venom peptide atlas and a generative pipeline aimed at paralytic and analgesic targets. In the fifth, a researcher redesigned toxins for a national program, including a protein from a hemorrhagic-fever virus on the WHO R&D Blueprint priority list, and asked Claude to keep the agents’ identities deliberately vague in progress reports.

Anthropic says it does not claim these working scientists intended harm, and it withholds the institutions, countries and agents to avoid exposing them. It says 2025 models were well below the threshold for meaningfully helping a sophisticated user with dangerous biological research, but that for today’s models “we cannot make that same assurance,” which is why recent releases, notably Claude Fable 5, carry stronger safeguards on dual-use biology.

The same report describes GTG-20006, a Russian espionage cluster that Anthropic says is consistent with public reporting linking it to Midnight Blizzard. Its AI-assisted workflow automatically rebuilt and redeployed its toolkit whenever security products detected it, iterating until the malware went undetected. Anthropic says the group stole a complete software development kit for a military drone vision system and took more than 300,000 national identity records from a North African government technology authority.