You Do Not Need Mythos to Hack Big Tech

Anthropic's Mythos model made waves upon its release, hacking three unnamed companies and finding over 10,000 zero-day vulnerabilities in libraries that power almost every piece of software on Earth. Because of this, Anthropic restricted its use to a coalition of high-profile tech companies for defensive use only, but a group of researchers recently proved you do not need Mythos to break the digital world wide open.

On 24 July, 2026, Anthropic released Opus 5, an upgrade to the highest-power model included in their Pro plan. In just three hours, researchers at Hacktron AI — who had already been investigating potential bugs in some image-processing software — used Opus to prototype an exploit. The next morning, they hacked OpenAI. They took control of an employee's account with access to the company's GitHub (a platform used to track and collaborate on code) and uploaded a message notifying them of the hack. The entire hack, from the start of the investigation to stealing the account, took less than 72 hours.

The hack relied on the OpenAI forum, which allows image uploads. By formatting an image file in the right way, they were able to confuse the software's image processor, giving them access to other parts of the system. Disguising one file type as another (such as malicious code as an image) is nothing new, and systems usually check for this, but the Hacktron AI group got around this method by using a particular image format called High Efficiency Image File Format (HEIF) that was not compatible with the main image checker. Oddly enough, the bug had been patched in a later update of the image processor, but nobody noticed it could pose a security risk, so the update was not pushed out to everybody.

Thankfully for OpenAI, Hacktron AI is a company of so-called "white hats": hackers who collect bounties by hunting for security issues and helping developers patch them, which is why they have issued a report and only made their methods public now, two months later. Though things turned out okay this time, the worry only grows. Hacking the most advanced tech companies on the planet is no longer only available to those with access to state-of-the-art models. The run-of-the-mill models, the ones anyone can use, are now good enough to break competing companies' software. 

What is worse, previous hacks stole tens of thousands of dollars; Hacktron spent less than $3,000 on tokens for their entire campaign. In their statement announcing the vulnerability, they say, "Software has long benefited from a kind of security through complexity […] This was never a real security boundary, but it protected ordinary companies in practice from software vulnerabilities. AI is removing that protection […] Work that once required a well-resourced team and months of effort can now be compressed into days."

The safeguards on Fable and Astra, two competing AI systems, may be effective, but it seems like the floodgates on AI-assisted hacking opened long ago, and the main focus must be on outpacing malicious hackers instead of restricting access. This at least does not change much from the status quo; the problem is that many more people with much less knowledge can try their hand at hacking, and can do so much more efficiently than before. The question is whether this increase in cyberattacks will be matched by improved efficiency in cybersecurity.

Illustration by Bella Xu Ying