Anthropic’s AI Misuse Report, Distilled Into 117 Findings

Earlier this month, Anthropic published a lengthy report cataloguing every case of Claude misuse it had detected. Daniel Meissler helpfully condensed the document into 117 distinct findings, and Bruce Schneier pulled out the highlights.

Dividing the labor between humans and AI

According to the report, AI agents are increasingly handling reconnaissance, exploitation, data theft, propaganda production, surveillance workflows, and research. Humans, meanwhile, still pick the targets, set the goals, and review the important outputs. The machine does the grunt work of an attack; the strategic calls stay with the operator.

Industrializing attacks

The report describes attackers using AI to industrialize credential theft, cloud compromise, phishing, vulnerability research, and the extraction of sensitive data from downstream organizations. Here AI acts as a scale multiplier, turning one-off operations into an assembly line.

Influence operations

Influence campaigns leaned on persistent agent memory, fake news sites, fabricated journalists, synthetic personas, political profiling, and large-scale multilingual content. One telling detail: for all that output, the high volume of content often produced little genuine engagement — quantity did not translate into reach.

Surveillance and repression

Surveillance and repression cases included automated dossiers, biometric and communications analysis, transnational targeting, and coercive recruitment. Notably, some systems kept operating locally even after model access had been revoked.

Dual-use risk

The biological and weapons cases illustrate the dual-use problem: AI supported advanced scientific and military work. Even so, the report generally does not establish any completed biological weapons or operational battlefield deployment. The potential is worrying, but no confirmed leap to real-world destruction has been recorded yet.

Source: Schneier on Security