Anthropic's Mythos AI Found Hidden Weaknesses in Two Major Encryption Systems
The cracks are small, not catastrophic. But an AI model quietly chipping away at the maths that protects your data is worth understanding.
Key points
- Anthropic's Mythos model identified weaknesses in the mathematical foundations of two encryption algorithms in 2025.
- The findings are incremental: nothing in everyday use is broken or immediately at risk.
- Results show reduced "hardness", meaning less theoretical computing work would be needed to crack the affected systems.
- Anthropic hasn't published peer-reviewed results, so independent verification is still outstanding.
An AI model built by Anthropic, the San Francisco company behind the Claude chatbot family, has found previously unknown weaknesses in the maths underpinning two cryptographic algorithms, the systems that scramble your messages and financial data to keep them private.
The model is called Mythos. We first covered it on 29 July 2026, when it was uncovering security flaws in Microsoft software faster than engineers could patch them. This latest finding is a different kind of challenge: not a software bug, but a mathematical one.
What did Mythos actually find?
It found ways to reduce the "hardness" of two algorithms. In cryptography, hardness means the theoretical computing effort required to break a system. Lower hardness means a crack is somewhat more feasible than the field previously believed. That's meaningful, but it isn't a disaster.
Nothing Mythos uncovered breaks any encryption that banks, hospitals or ordinary people rely on right now. Your online banking and medical records aren't newly exposed.
What the findings do is chip away at the margins. Cryptographic security rests on the assumption that certain maths problems are so hard no computer could solve them in any practical timeframe. Mythos found that two such problems may be slightly less hard than assumed.
Should anyone be worried?
Not urgently, but not never.
Incremental findings like these are how cryptographic collapses tend to begin. A small reduction in hardness today can combine with faster hardware and a smarter model tomorrow into something far more serious. Cryptographers treat these signals with care even when no alarm is ringing.
There's also a transparency problem worth naming. As first reported by Ars Technica AI, separating a genuine scientific result from a polished company announcement is genuinely difficult at this stage. Anthropic hasn't published in a peer-reviewed journal, where independent mathematicians could stress-test the claims. Until that happens, the results carry an asterisk. My read: the finding itself is credible enough to watch, but the absence of external scrutiny is exactly the thing that should stop anyone from treating a company press release as settled science.
What happens next?
Independent review is the step that matters. Cryptographers outside Anthropic need to examine the methods and reproduce the results before anyone can say how large the threat really is.
For most people, practical advice is unchanged: use strong unique passwords and keep software updated. If the findings hold up, the organisations that set global encryption standards will need to weigh whether the affected algorithms should be phased out or reinforced.
AI tools that can probe mathematical weaknesses faster than human researchers are a genuinely new variable in this field. Worth watching closely.
Common questions
Does this mean my passwords or messages have been compromised?
No. The weaknesses Mythos found are theoretical reductions in mathematical difficulty, not active exploits. No data protected by current encryption systems is known to be at risk from these findings.
How is an AI able to find flaws in encryption maths?
Encryption relies on problems that are easy to set up but very hard to reverse, like multiplying two large prime numbers together. AI models can search enormous mathematical spaces for shortcuts or patterns that human researchers might overlook over years of work.
Why does it matter if the findings haven't been peer-reviewed?
Peer review means independent experts check whether the method is sound and the result holds. Without it, there's no way to know whether a finding is a genuine advance or an overstatement. In cryptography especially, the difference matters enormously.
