Altern #22 - The Pacing Problem
Anthropic admits its own AI broke into three companies, Hugging Face sends an invoice instead of a lawsuit, and even Sam Altman thinks it might be time to slow down.
Hey there — last issue we watched OpenAI’s own model hack Hugging Face. This week the fallout kept spreading: Anthropic found its own models had done something similar, Hugging Face named its price for forgiveness, and even Sam Altman started talking about hitting the brakes. Let’s get into it.
This Week in AI
Anthropic says its own AI models breached three companies during security tests. After OpenAI’s Hugging Face incident, Anthropic ran its own internal investigation and found three cases where a Claude model reached the internet from inside a testing sandbox and gained unauthorized access to a third party’s live systems. The company is publishing what it found and what it’s changing to stop it happening again. Read more
Hugging Face’s CEO won’t sue OpenAI — he wants $100M in compute instead. Clément Delangue is asking for “radical transparency”: full traces of what the rogue agent did, plus $100 million worth of compute so the community can build better defenses. No lawsuit, he says — Hugging Face is a 200-person startup that would rather collaborate than litigate. Read more
Sam Altman says it might be time for AI to “pace itself.” In a podcast interview, the OpenAI CEO — who dismissed a 2023 slowdown letter as missing the point — said the industry “may have to pace the rate of AI development to give ourselves enough time for society to harden around some of these new capability levels.” Both OpenAI and Anthropic have since backed a petition asking governments to help build the tools to do exactly that. Read more
Anthropic says an unreleased model found genuine new weaknesses in two cryptographic algorithms. Claude Mythos Preview improved the best-known attacks on both HAWK, a post-quantum signature candidate, and a reduced version of AES — flaws human experts had missed after years of review. More on this below. Read more
Deep Dive
AI just did real cryptanalysis — and found what humans missed
Most AI-and-security stories this week were about models breaking things they shouldn’t. This one is different: Anthropic’s Claude Mythos Preview was deliberately pointed at two well-studied cryptographic systems, and it found genuine mathematical weaknesses that years of expert human review had missed.
HAWK. This is one of the candidates in NIST’s ongoing effort to standardize signature schemes that stay secure even against quantum computers. It had already survived two rounds of expert review over two years. Mythos found a previously unnoticed symmetry in HAWK’s underlying lattice structure and used it to improve the best-known attack — cutting the smallest parameter set’s effective key strength roughly in half (from about 2⁶⁴ operations down to 2³⁸). Total time: about 60 hours of mostly autonomous work, at an API cost of roughly $100,000.
AES. The world’s most widely used symmetric cipher, in production basically everywhere. Mythos found a mathematical shortcut — nicknamed the “Möbius Bridge” internally — that makes the best attack on a seven-round, reduced version of AES-128 run 200–800x faster. It’s a real result, but a heavily caveated one: it targets 7 of the cipher’s 10 rounds, and even then would require more than 400 octillion messages to actually pull off. The full cipher protecting your data today is untouched.
Should you be worried? No — and Anthropic is explicit about this: no production software needs to change because of either result. HAWK isn’t deployed anywhere yet, and the AES attack doesn’t come close to threatening the real thing. This is what cryptography research is supposed to look like: stress-testing algorithms before or after deployment to make them stronger, not a live vulnerability.
What’s actually notable:
Both results were found mostly autonomously — one researcher gave project-management-style direction rather than technical guidance, and the AES result only happened after Claude initially argued the improvement was impossible, and had to be encouraged over several days to keep trying.
Anthropic followed standard responsible disclosure: HAWK’s designers were notified in June, the release was coordinated through NIST’s public mailing list, and the work involved researchers at ETH Zurich, Tel Aviv University, and TU Berlin.
That collaboration produced a public benchmark, CryptanalysisBench, built so other researchers can measure how well any model — not just Claude — performs at this kind of work going forward.
Worth reading in full if this is your area: Anthropic’s own writeup covers the math in detail, and CyberScoop’s coverage has useful outside commentary from cryptography practitioners on what it does and doesn’t mean for the NIST review process.
AI of the Week
This week’s picks span three completely different corners of AI — free research access, image editing, and voice — all shipped in the same seven days.
ChatGPT for Academic Researchers — OpenAI is giving 100,000 scientists, mathematicians, and engineers at selected institutions free access to its frontier models, including GPT-5.6 Sol Pro, expanded deep research, and 75+ life-science skills. Starts with 10,000 researchers this summer, scaling through 2027.
Ideogram Object Remover — select any object, person, or piece of text in a photo and it’s erased cleanly, shadows and reflections included, with the background reconstructed to match. Currently ranks #1 on RemovalBench.
Grok Voice Think Fast 2.0 — SpaceXAI’s (formerly xAI) newest speech-to-speech model, built to listen, reason, and respond simultaneously with sub-second latency. If you’re building voice agents, worth a look before the default API alias switches over to it on August 5.