That artificial intelligence (AI) exposed a weakness in a post-quantum cryptography candidate doesn’t mean that encryption has failed, but that security must be designed to withstand failure.
On 28 July 2026, Anthropic disclosed that Claude Mythos Preview had identified a serious structural weakness in HAWK, a post-quantum digital signature scheme created by some of the world's sharpest cyber security minds to withstand future quantum computers.
The alarm sounded worldwide because Mythos Preview did something IT leaders had anticipated, but not quite like this. Instead of a quantum computer, it was AI that struck first – reducing the estimated effort needed to crack HAWK's smallest configuration by about 67 million times.
The proposed mitigations – including substantially increasing parameters or changing the underlying construction – would have made HAWK too large or slow to remain competitive. The following day, its developers withdrew their candidate from the National Institute of Standards and Technology’s (NIST’s) post-quantum digital-signature standardisation process.
Separate research raised questions about another candidate, SQIsign. The three post-quantum standards NIST finalised in 2024 are unaffected – but the episode reinforces why no algorithm, however carefully designed, can be considered permanently secure.
The cracks appear
On NIST’s public forum, researcher Cong Ling described the discovery as historic, pointing out that humans make mistakes and are slower than AI. We are, he said, entering a new era – Post-AI Cryptography.
This came amid a bid by the tech sector to deploy AI to find and repair vulnerabilities before criminals could exploit them.
In April, Anthropic launched Project Glasswing with founding partners including Microsoft, Google, AWS, Apple and Cisco, later expanding to around 200 organisations, including NATO and the European Union’s ENISA. Within weeks, it had uncovered more than 10 000 high- or critical-severity vulnerabilities.
For decades, we have treated encryption as a mathematical fortress, believing sufficiently complicated calculations would keep information safe.
Defensive AI is also moving into commercial application with Palo Alto Networks’ Unit 42 Continuous Frontier AI Defense, using commercial and open-weight AI models to examine applications, application programming interfaces and cloud infrastructure, identify vulnerabilities and recommend repairs.
Fire with fire
This doesn’t mean AI can automatically repair cryptographic mathematics, but it demonstrates how industry is increasingly using machines to find security holes before attackers exploit them.
The fightback is no longer just about developing stronger algorithms, as it has been for thousands of years dating back to the ancient Egyptians’ use of mathematical techniques to conceal messages, but increasingly about using machines to fight machines.
Parallel to this, concerns are growing over AI models misbehaving. Microsoft AI chief Mustafa Suleyman has warned that giving AI humanlike characteristics increases the risk of systems going rogue. Meanwhile, Anthropic CEO Dario Amodei has called for regulation and slower development, while Meta CEO Mark Zuckerberg favours independent evaluators.
This isn’t about whether AI can discover weaknesses faster than humans, but whether researchers can verify them quickly enough. Some are calling for independent expert review and reproducible code, with Anthropic acknowledging that human researchers could become a bottleneck.
The deeper problem
But there is a fundamental problem. Despite advances in AI-assisted cyber security and post-quantum cryptography, we are blinkered like horses; focusing on building stronger defences to keep attackers out. We develop algorithms, identify weaknesses, repair them and repeat the process.
For decades, we have treated encryption as a mathematical fortress, believing sufficiently complicated calculations would keep information safe. Yet supposedly secure systems have a habit of eventually being broken.
In 1994, Peter Shor demonstrated how a sufficiently powerful quantum computer could undermine widely used public-key cryptography. We have spent more than three decades preparing for that possibility, only for AI to expose a weakness in an algorithm designed to withstand precisely that threat.
While Anthropic’s discovery was the wake-up call, this doesn’t mean cryptography has failed, but that we must stop treating security as a target to be achieved and rather reconsider our mindsets.
Same warning, new weapon
As an industry, we need to shift to a more philosophical and holistic approach that enables us to create systems capable of adapting, containing damage and recovering when their protections fail. This breaks what would otherwise be an endless loop of design and break.
Cryptographic agility, defence in depth and resilience are established principles, but they can no longer be secondary considerations or something to worry about in the 2030s. The mathematics protecting one generation can be undermined by the computing capabilities of another, and security must be designed around that reality.
More than eight decades after Turing helped develop the Bombe to crack Enigma, machines are once again changing the rules of cryptography. Only this time, they must become part of the defence as well as the attack.
The question is no longer whether our digital locks can be broken, but how we design security that can survive when they are. Turing taught machines to break codes and now we must teach them to protect what they can break.
Share
