Mentatcurated
Artificial Intelligence high · independent

The slow part is us

Claude took three days to find a faster attack on a deliberately weakened version of AES. Two human cryptographers then took nearly a month to satisfy themselves it was correct.

Claude spent about three days on a stripped-down version of AES, the cipher that protects most internet traffic, and came back with an attack 200 to 800 times faster than the best one in the literature. Two human cryptographers then spent nearly a month checking that it held up.

"There's nothing easy to find; this is the most-studied block cipher in existence." — Claude, before it found something

That was one of two findings Anthropic's red team published. In the other, the same model worked semi-autonomously for roughly sixty hours on HAWK, a candidate signature scheme meant to survive quantum computers, noticed a symmetry in its underlying lattice that nobody had, and cut the work of recovering a key by a factor of about seventy million. A companion paper from ETH Zurich, Tel Aviv University, Haifa and Anthropic supplies the systematic version. CryptanalysisBench wires up 191 schemes from four NIST standardization contests so that an agent has to win the formal security game against a working implementation, not describe an attack. Two different models independently derived the same complete key recovery against one lightweight scheme from just two queries, a flaw nobody had reported, and one found an error in a published security proof.

Nothing you use is weaker this morning. HAWK is deployed nowhere, the AES attack is on a research variant with three of its ten rounds removed, and of fourteen full-strength breaks on the benchmark only two exposed a genuine design flaw; the other twelve were bugs in someone's implementation. Each of the two headline findings cost around $100,000 in model calls.

What is moving is the economics of scrutiny. A cipher earns trust by surviving years of attention from the few hundred people qualified to attack it, which is why NIST competitions run for most of a decade, and machine attempts now cost less than that attention. The benchmark also answers the bottleneck its companion exposes: winning a security game against real code is something a computer can confirm, so the month two experts spent checking Claude's arithmetic is exactly the step a harness like this removes.

The lenses

Novelty 4
Impact · breadth 3
Impact · depth 4
Actionable 3
Delight 4
Substance 5
Hype 3

The facts

The benchmark191 schemes from four NIST competitions, published with code
Who built itETH Zurich, Tel Aviv University, University of Haifa and Anthropic
CostRoughly $100,000 in model calls per finding
Risk to deployed systemsNone: the quantum-resistant scheme is unused, full AES is untouched
Open arxiv.org →

How this connects

Tap a node to open it

How this connects