Saturday, 7 November 2026
DeepMind's AI Agents Find Two Weaknesses in Old 1024-Bit RSA Encryption
DeepMind’s AI Agents Find Two Weaknesses in Old 1024-Bit RSA Encryption
Google DeepMind said on Friday that AI agents, guided by its researchers, found two weaknesses in legacy 1024-bit RSA. The first is about digital signatures: the “padding” added before signing gives no real protection once an attacker controls roughly half of the signed block. In a sandbox, on a 1024-bit benchmark key that academics had already broken, the agents reproduced a forgery with 3.5 times less CPU time; for the harder case of a short 32-byte fingerprint, a new algorithm is, by DeepMind’s own theoretical estimate, about 1,600 times better than the best known method. The second is a batch trick: factoring many 1024-bit keys together now costs about 30 times less per extra key. The work was led by Dj Dvijotham and Abhradeep Thakurta with cryptographers Nadia Heninger and Emmanuel Thomé; no live system was touched, and 2048-bit keys, the modern standard, are not affected. Cryptanalysis used to move in decades — the agents just shortened the warranty on anything still running the old keys.
Source: Pushmeet Kohli, Google DeepMind, on X, November 6, 2026
Anthropic Studies 207,049 Research Sessions: AI Could Take On a Median 38% of a Scientist’s Time
Anthropic published a report on Thursday on how scientists use Claude, built on 207,049 research sessions from 2 August to 30 September, plus a survey of 528 scientists. Computer science leads (11.9% of sessions on Claude, 32.7% on Claude Code), followed by engineering and medicine. With today’s models, widely adopted, AI could take on 27-52% of the researcher time a new study needs, a median of 38%. If 75-100% of the non-physical steps were automated, fields like statistics and computer science could reach about 15 times today’s output; biology and chemistry only 3-4 times, because the lab bench doesn’t speed up. About 40% of the surveyed scientists expect AI to run whole research projects by 2030, 64% by 2035, and 16% say never — with the caveat that they were recruited from Anthropic’s own science-grant applicants, not a random sample. The bottleneck in science, it turns out, is still a pipette.
Source: Anthropic, November 5, 2026
Five AI Firms Agree on a Shared Scale for Rating Cyber Jailbreaks
The Frontier Model Forum published version 1.0 of a Jailbreak Rating Framework for Frontier Cyber Threats, backed by Amazon Web Services, Anthropic, Google DeepMind, Microsoft and OpenAI. The point is a common yardstick: when someone gets past a model’s safeguards, how much dangerous cyber capability does that unlock, and how easily? The rating combines four scores: capability gain, capability scope, ease of deployment and ease of discovery. The Forum calls it preliminary, will test whether different assessors reach the same rating, and takes public comments until 21 December. OpenAI announced its part on Friday. For an industry that grades its own models on its own scales, agreeing on one ruler for the break-ins is a start.