Le prompt, toujours une faille majeure des LLM

A study using the HackAgent framework performed extensive automated jailbreak testing against Opus 4.8 and Fable 5, finding static obfuscation largely ineffective but adaptive iterative attacks (Tree-of-Attacks) able to bypass safeguards (11.5% success on Opus, up to 6.1% on Fable) and produce dangerous social-engineering outputs like phishing and ransom notes; however, Fable 5 consistently refused to generate exploitable malware or exploits, and the report recommends defenders move from keyword filters to semantic, multi-turn monitoring.