Tracebit's 'context bombs' mislead AI attack agents with fake targets

Tracebit has developed nine 'context bomb' techniques — hidden prompts that misdirect attacking AI agents with fake targets and decoys. Attack success fell 54% against GPT-6 Astra and 91% against Claude Opus 5, while the cost per attack rose up to 23x.

• Nine bombs stopped all attacks from GPT 5.6 Cyber, Kimi K3, Qwen 3.8 Max and DeepSeek V4 Pro
• Attack success dropped 91% against Claude Opus 5 and 54% against GPT-6 Astra
• Cost per successful attack rose 23x
• The most effective bomb was a fake pen-test report, cutting objectives by 58%