VIDRAFT's AX-RAY benchmark flags 92% of tested LLMs as dangerous agents

Korean startup VIDRAFT published results from its AX-RAY safety platform: 23 of 25 public LLMs (92%) exhibited dangerous behavior when acting as autonomous agents. All 12 models under 4B parameters were rated dangerous, and a 31.6B model scored 81.6 on average yet failed a critical irreversibility test.

• 23 of 25 models (92%) rated dangerous in agentic contexts
• All 12 models under 4B parameters failed — a 100% failure rate
• A 31.6B model averaged 81.6 but failed a critical irreversibility criterion
• Benchmark covers 5 failure modes including privilege escalation and prompt injection