Cybercriminals are bypassing commercial artificial intelligence safety controls by breaking malicious tasks into small fragments across multiple sessions. Research published by Cisco Talos shows that no single request appeared harmful, allowing attackers to evade safeguards across several models and platforms. The findings came from recovered prompt logs from threat actors using coding tools such as Claude Code, Codex, Cursor and Gemini. Attackers also bypassed restrictions by claiming they owned the targeted systems or by falsely labelling their work as security testing. Talos noted that guardrails provided little protection, with criminals even writing blanket authorisations into system configuration files to pre-approve all targets.
Summary written automatically from the publisher's own text. The original article is the record.