Anthropic released its third report on AI misuse since March 2025, detailing several interceptd attempts at malicious activity. the company reported blocking efforts ranging from cyberattacks to research that could potentially facilitate biological weapons.
The intercepted chikungunya virus research request
Anthropic identified a specific instance where its Claude models were used to attempt to draft a grant application for scientific funding. According to the report, this application involved gain-of-function research on the chikungunya virus, a mosquito-borne pathogen that causes severe fever and pain. The research in question specifically targeted the virus's ability to evade the immune system and increase its transmissibility .
While such research is vital for vaccine development, Anthropic warned that these capabilities could also be used to make a pathogen more dangerous. The company has since implemented stronger safeguards in its latest models to restrict biological research that could be repurposed for weaponry.
Nine influence operations across the Gulf and Europe
The company's findings also included the detection of nine distinct cases of coordinated social media manipulation. These operations involved the creation of hundreds of fake accounts designed to mimic ordinary people while amplifying specific political viewpoints. As reported by Reuters, these activities were detected originating from various regions, including South Asia, Africa, the Gulf, and Europe.
Anthropic's Chief of Staff, Elizabeth Gribkoff,noted that identifying these operations while they are still being built allows for the potential prevention of disinformation before it spreads. The company aims to shine a light on these threats to prompt both governments and AI competitors to take similar defensive actions.
Chris Olah’s resignation and the IPO tension
The release of this report follows the immediate resignation of co-founder and leading researcher Chris Olah. In a public statement, Olah argued that both Anthropic and the broader AI industry are failing to prioritize risk mitigation, specifically regarding the danger of AI systems eluding human control.
This internal conflict arises at a sensitive time for the organization, as the AI startup is currently planning an initial public offering for this coming fall. The tension highlights a growing debate within the industry regarding whether commercial speed is being prioritized over safety protocols.
The gap between Claude Opus 4 and Claude Fable 5 safeguards
Anthropic highlighted a significant shift in safety requirements as its models transition from older versions to more advanced iterations. The company stated that older models, such as Claude Opus 4 and Claude Sonnet 4.5, were not capable of assisting in sophisticated biological research. consequently, the safeguards on those models were less stringent, focusing primarily on preventing novices from recreating known bioweapons.
However, for newer models like Claude Fable 5, Anthropic has implemented much stronger restrictions on "dual-use" biological research queries. One lingering question involves the "illicit distillation" case mentioned in the report; it remains unclear how these industrial-scale capaigns successfully extract model capabilities without authorization. Furthermore, the report does not name the specific state-sponsored groups involved in the propaganda campaigns, leaving the exact nature of the actors unverified.
Comments 0