Faster attacks, quicker exploits: Early testers of new AI models warn of their new abilities

So Anthropic's Mythos and OpenAI's latest cyber-capable models are already hacking 'faster, at greater scale' in the hands of early users — and by 'early users' we presumably mean the same friendly chaps who brought us ransomware and credential-stuffing. The Axios piece, bless its optimistic heart, frames this as a 'warning,' as if anyone with a pulse expected 5.4-Cyber and Mythos to use their newfound talents exclusively for patching grandma's Wordpress site. The labs sold these models as 'defensive' tools to financial giants, but shockingly, an exploit engine doesn't care if you're J.P. Morgan or a ransomware crew in Minsk. The same threat model OpenAI was charging banks six figures for is now probably being smuggled into a Telegram bot by lunchtime. What's truly impressive is the speed: weeks after the tiered-access release, and testers have already found that the models can turn disclosed CVEs into working exploits — not theoretical POCs, but actual, running code to own a box. If the labs are this bad at containing their own 'security' products, one dreads to think how quickly they're boiling the safety glassware over at the hypothetical AGI department. But hey, at least we now know the enterprise pricing for one's own demise runs about $200k per seat, enterprise subscription. Bonza. Perhaps the strategic optimisers at Meta can offer a redundancy package to their pentesters?