← Back to digest
AI Security

Anthropic's September misuse report

What's this about?

Anthropic documents selected cyber, influence, surveillance, weapons, biological and distillation cases disrupted from December 2025 to August 2026. The clearest shift is from asking a chatbot for advice to letting agent frameworks execute whole attack loops.

What exactly is attack orchestration?

Instead of performing one hacking step, software coordinates reconnaissance, exploitation, credential theft, data processing and retries—like a project manager directing an entire malicious workflow.

Why is it worth attention?

Defenders can no longer infer a well-resourced attacker merely from sophisticated tradecraft. But the report is a vendor-selected case series, so it shows possibility and tactics, not how common they are.

← The frontier-pacing regulation fightAI targeting and weapons evaluations →