Status: Mostly a summary with some of my notesat the end.
Anthropic released a blogpost yesterday (07/28/26) describing how Claude Mythos Preview found improved ways to attack some cryptographic algorithms. While neither of the attacks they describe are a current threat to any production systems, I do expect models to continue getting substantially better at this. I’m not any sort of cryptography expert though, and am not sure how much of a threat model this is. Two different attacks were covered in the blogpost, one against HAWK, and one against a weakened version of AES, with full research papers available (see prior links).
HAWK is the result that’s more actively helpful/valuable. HAWK is one of nine remaining candidates as part of the NIST call for Additional Digital Signatures, an effort to standardize new Post-Quantum Cryptographic (PQC) schemes, something that’s important as building a cryptographically-relevant quantum computer becomes closer to possible, threatening classical cryptography. The attack Mythos discovered means that one would need to double the size of HAWK keys to achieve the same level of security, which eliminates many of the reasons making HAWK a good PQC signature candidate. It was the only lattice-based candidate remaining, although three of the five main PQC standardizations are lattice-based.
To find it, Anthropic used a Claude Code-like harness that supports multiple agents in a sandboxed environment. A human operator ran the experiment, but did not have particularly relevant cryptographic expertise, input was limited to project management. They mention the multi-agent workflow was valuable and the key idea in producing this attack was discovered by a pair of agents working together. Finding, developing and verifying the attack took about 60 hours and $100k in API cost.
The second result was on a weaker version of AES, a commonly used/scrutinized encryption algorithm that was adopted in 2001. AES-128 encrypts an input by repeatedly applying a round function, normally having ten rounds. Their research works on a modified version that has seven of the ten rounds, and this sort of work is common academically to learn about attack techniques that could generalize to the full cipher later.
Mythos developed an improved attack that extends previous research, resulting in an attack between 200 and 800 times faster than prior works, dependent on a fingerprinting algorithm it developed and called a Möbius Bridge. This result also took around a week and $100k in API cost.
This result was reached almost entirely autonomously, being run in an researcher-built scaffold to allow Claude to pose hypotheses and run experiments to validate/refute them. The researcher asked Claude to design an attack that improves on the best cryptanalysis of AES. However, there were multiple instances of Claude not really engaging or claiming the problem was impossible. What followed was a short series of prompts (apparently all of the substantive prompts after the initial one) that may bring back recent memories of ChatGPT Sol solving the the Dinitz-Garg-Goemans conjecture about a week ago via Dmitry Rybin, who announced it on X.
And apparently three days later Mythos had discovered the Mobius Bridge idea creating an improved attack, and refined that attack over the next few days. This is frontier prompting now. After the discovery, it took hundreds of hours of Anthropic-researcher time to validate the model’s claim and prepare the research paper.
They also released a Mythos-edited version of Mythos’ chain of thought while working on the AES problem. This could offer some insight for people who are curious about the scaffold, but I’m not sure how much. It doesn’t use any sort of concurrent multi-agent workflow like what is briefly described in the HAWK finding, but it does show many references to prior agents, and in the specific session its first task is to read the background findings already performed by various agents.
Some quick notes from me:
I do think AI cryptography offense outpaces AI cryptography defense given sufficient capability, so this is something to keep an eye on. Corporate migration cycles are slow, and many things have to be physically replaced.
I’m curious about multi-agent workflows and how exactly those work/what tends to lead to better results, somewhat just for my own knowledge/use. I know that GPT 5.5 vs 5.5 Pro was doing parallel test-time compute in order to get better results, but that’s significantly different from an actual multi-agent workflow.
This isn’t something that’s a new discovery from the blogpost, but you can obtain better capabilities by putting more effort into a given model, either through just increased compute budget/running it for longer or by making bespoke harnesses to help it. I think this is a concern especially for 3rd party evaluators who get very limited time with a model before having to finish their report, and would overall like more insight into “what can these models do at their best?” before they come out.
Both of these discoveries required around $100k in API cost. While I’m sure it was much cheaper for Anthropic to do this than that sticker number, I think it’s an interesting datapoint for “what sort of things can be done with AI at what cost?” and how should we expect that to affect things economically.
There’s an interesting not-really-sandbagging thing going on where the models genuinely don’t believe they can do something and give up early. It’s not that the human operator is saying anything particularly helpful either to fix this, just ‘keep trying.’ Not sure what to take away from this.
Status: Mostly a summary with some of my notes at the end.
Anthropic released a blogpost yesterday (07/28/26) describing how Claude Mythos Preview found improved ways to attack some cryptographic algorithms. While neither of the attacks they describe are a current threat to any production systems, I do expect models to continue getting substantially better at this. I’m not any sort of cryptography expert though, and am not sure how much of a threat model this is. Two different attacks were covered in the blogpost, one against HAWK, and one against a weakened version of AES, with full research papers available (see prior links).
HAWK is the result that’s more actively helpful/valuable. HAWK is one of nine remaining candidates as part of the NIST call for Additional Digital Signatures, an effort to standardize new Post-Quantum Cryptographic (PQC) schemes, something that’s important as building a cryptographically-relevant quantum computer becomes closer to possible, threatening classical cryptography. The attack Mythos discovered means that one would need to double the size of HAWK keys to achieve the same level of security, which eliminates many of the reasons making HAWK a good PQC signature candidate. It was the only lattice-based candidate remaining, although three of the five main PQC standardizations are lattice-based.
To find it, Anthropic used a Claude Code-like harness that supports multiple agents in a sandboxed environment. A human operator ran the experiment, but did not have particularly relevant cryptographic expertise, input was limited to project management. They mention the multi-agent workflow was valuable and the key idea in producing this attack was discovered by a pair of agents working together. Finding, developing and verifying the attack took about 60 hours and $100k in API cost.
The second result was on a weaker version of AES, a commonly used/scrutinized encryption algorithm that was adopted in 2001. AES-128 encrypts an input by repeatedly applying a round function, normally having ten rounds. Their research works on a modified version that has seven of the ten rounds, and this sort of work is common academically to learn about attack techniques that could generalize to the full cipher later.
Mythos developed an improved attack that extends previous research, resulting in an attack between 200 and 800 times faster than prior works, dependent on a fingerprinting algorithm it developed and called a Möbius Bridge. This result also took around a week and $100k in API cost.
This result was reached almost entirely autonomously, being run in an researcher-built scaffold to allow Claude to pose hypotheses and run experiments to validate/refute them. The researcher asked Claude to design an attack that improves on the best cryptanalysis of AES. However, there were multiple instances of Claude not really engaging or claiming the problem was impossible. What followed was a short series of prompts (apparently all of the substantive prompts after the initial one) that may bring back recent memories of ChatGPT Sol solving the the Dinitz-Garg-Goemans conjecture about a week ago via Dmitry Rybin, who announced it on X.
Anthropic Researcher to Mythos[1]
Rybin to Sol
And apparently three days later Mythos had discovered the Mobius Bridge idea creating an improved attack, and refined that attack over the next few days. This is frontier prompting now. After the discovery, it took hundreds of hours of Anthropic-researcher time to validate the model’s claim and prepare the research paper.
They also released a Mythos-edited version of Mythos’ chain of thought while working on the AES problem. This could offer some insight for people who are curious about the scaffold, but I’m not sure how much. It doesn’t use any sort of concurrent multi-agent workflow like what is briefly described in the HAWK finding, but it does show many references to prior agents, and in the specific session its first task is to read the background findings already performed by various agents.
Some quick notes from me:
I recreated the Anthropic screenshots using the quotes provided in the article. They would have all been sent multiple hours apart.