Beyond the AI buzz words, the doom and gloom of job losses, there is a much darker side.
Anthropic‘s decision to publicly release its latest frontier AI model, Fable 5 (nee Mythos), has reignited familiar debates about whether powerful AI systems should be made broadly available. In announcing the model, the company acknowledged that its capabilities, particularly in cybersecurity, could potentially be misused to cause serious harm and outlined a range of safeguards designed to prevent abuse.
The discussion that followed has largely centred on whether those safeguards are sufficient.
That may be the wrong question.
According to Charles Guillemet, Chief Technology Officer at Ledger, attackers are unlikely to be constrained for long by the safety mechanisms embedded within large language models. History suggests that guardrails are often bypassed through prompt engineering, jailbreaking techniques or simple persistence.
“If you’re reassured that Anthropic has only shipped a ‘safe’ version of Mythos, don’t be,” Guillemet argues. “Large language model safeguards have repeatedly shown they don’t survive contact with even the laziest adversary.”
The concern is not merely theoretical. Cybercriminals already have access to a growing ecosystem of AI-powered tools, and the volume of cyberattacks worldwide continues to rise. Whether or not Fable 5 itself is perfectly aligned may ultimately matter less than the broader reality that sophisticated offensive capabilities are becoming cheaper, faster and more accessible.
That shifts the burden from restricting models to strengthening infrastructure.
For years, cybersecurity has relied heavily on patching vulnerabilities after they are discovered. The emergence of highly capable AI systems accelerates the pace at which weaknesses can be identified and exploited, exposing the limitations of reactive security approaches.
Guillemet argues that genuine resilience will come from security-by-design principles, including formal verification and hardware-based secure enclaves that minimise the impact of software vulnerabilities. In other words, organisations must assume that attackers will eventually gain access to advanced AI capabilities and build systems that remain secure regardless.
This is where the real challenge lies. Despite growing awareness of cyber risk, many organisations continue to delay software updates, maintain legacy systems and underinvest in foundational security architecture. As AI lowers the barriers to offensive cyber activity, those weaknesses become increasingly difficult to justify.
The release of Fable 5 is therefore less a story about one model and more a warning about the future of cybersecurity. The question is not whether AI can help attackers. It already does. The question is whether governments, businesses and individuals are prepared for a world in which those capabilities are available to everyone.
The answer, at present, remains uncertain.
