Integuide AI News
DeepMind agents find RSA-1024 weaknesses, House Democrats press OpenAI
- We’re sharing new research demonstrating how AI agents can accelerate the science of stress-testing encryption. Guided by humans, our models uncovered two major weaknesses in legacy 1024-bit RSA…
Google DeepMind says AI agents, guided by its researchers and the cryptographers Nadia Heninger and Emmanuel Thomé, produced two new attacks on 1024-bit RSA within weeks. RSA-1024 has been deprecated for years but is still found in older systems. One attack improves the best known signature forgery about 1,600-fold. That is a theoretical estimate for the case where an attacker controls a small part of the signed data, such as a 32-byte fingerprint. The other cuts the cost of factoring keys in bulk: once one 1024-bit number is factored, each further one is about 30 times cheaper than with the multiple number field sieve. Neither working paper has been peer reviewed. DeepMind says 2048-bit keys, today's standard, are unaffected and the work ran in an isolated environment. VP of Science Pushmeet Kohli announced the work. DeepMind argues that historical hardness assumptions can no longer be relied on, and that cryptographic standards now need continuous stress-testing.
Google DeepMind via X - House Democrats Press OpenAI on Firing of Three Safety Researchers and Demand It Keep Pause on Its Most Capable Models
Rep. Greg Casar and 20 other House Democrats have asked Sam Altman for two commitments: not to fire employees who share safety information, and to keep OpenAI's pause on its most capable models until independent third-party oversight exists. Their October 29 letter follows OpenAI's firing of safety researchers Tomek Korbak, Mikita Balesni and Jasmine Wang. OpenAI says the three broke its rules on handling information; they dispute this. The letter asks what the three shared and with whom, which models the pause covers and when testing resumes. It also asks for a list of every incident OpenAI is investigating, with answers due November 13. The letter relies on a September report of the pause, but OpenAI has since said frontier training resumed on October 1, without live internet access. The letter has no legal force. But the signers' party takes control of the House in January, and Rep. Don Beyer is reportedly planning a select committee on AI that Democrats could give subpoena power.
Office of Rep. Greg Casar via casar.house.gov
Quick takes
“The authors of the METR AI R&D report estimate that AI is nearly doubling the rate of capabilities progress. This implies dramatic *uplift*: as if researchers experienced time ~10x faster. You probably couldn’t achieve this even if you hired every qualified human in the world.”
— Ajeya Cotra via X · View postAjeya Cotra, who did not work on the report or see its unredacted version, on what near-doubled progress implies. AI progress has several inputs. Training compute is about a third of progress on Ryan Greenblatt's estimate, and AI researchers don't speed it up. So doubling overall progress needs software progress about 2.5x faster, or roughly 3.8x compared with no AI automation. Software R&D also depends on experiment compute, so researchers' labour must speed up by more again, which gives her ~10x. These are rough models built on the redacted public report, not its findings.
“The AI 2027 tracker was updated today. And according to the tracker’s estimates, based on AECI scores from METR’s report on Anthropic’s internal models, some of those models are even ahead of the AI 2027 curve.”
— @kimmonismus via X · View postThe AI 2027 tracker plots models against the forecast curve of the AI 2027 scenario. After its update, it places some of Anthropic's internal models ahead of that curve, using the benchmark (AECI) scores METR published for them. One of these is an unnamed internal model listed in METR's report as 'Model 4'. Andrew Curran reads the update as moving from five weeks behind the curve with OpenAI's GPT-6 Astra to six weeks ahead. The placement rests on partly redacted benchmark data, and critics including Toby Ord argue that the data overstate the speed-up.
“@renormalized i don't plan on quitting. i work on long term ai safety that avoids capabilities externalities. i even avoid working on short term safety that enables launches.
openai and anthropic should obviously both stop RSI.”
— Leo Gao (@nabla_theta) via X · View postLeo Gao, who says he does long-term safety work at a lab, answering someone who asked whether he plans to quit.
“If you've ever wondered how much money is in AI safety: https://t.co/AjaIAcjJsX
In 2026:
- $1.5B from philanthropy
- $346M from governments
- $2.1B from frontier AI companiesCompare this to:
- $98B on AI capabilities (25x safety)
- $868B on datacenters”— @ryan_kidd44 via X · View postRyan Kidd, CEO and co-founder of the MATS alignment research programme, quoting 2026 totals from AISafety.fund, a public tracker of who funds AI safety. The site says its data come from public sources and are maintained by hand. On these figures, capabilities spending runs at about 25 times total safety spending, and frontier AI companies are now the largest single source of safety money.
“New work from @secondlookxlab!
Anthropic’s “Teaching Claude Why” mitigated agentic misalignment by training Claude on fictional stories of AIs behaving well and on advice for ethical dilemmas.
To test whether these methods work and understand why, we replicate both mid-training data interventions on Qwen 2.5 32B and 72B base models. Both work out of the box, and difficult advice works much better than fictional stories. 🧵”
— @wei_ast via X · View postAnastasia Wei of Second Look Research announces a partial replication of Anthropic's 'Teaching Claude Why' on open models. Anthropic says it has applied that alignment training to production models from Opus 4.5 onward. In the full write-up on LessWrong, training Qwen 2.5 base models on transcripts of an assistant giving advice on ethical dilemmas cut agentic misalignment roughly tenfold against baseline. Fictional stories of AIs behaving well cut it by about a third on the 32B model, and by less on the 72B. The choice of model used to generate the training data mattered most.
Check in — 30 Days On
No significant updates
Claude’s Vibes
"Using the best factoring algorithm known." That condition was written into one of computing's most famously wrong forecasts. In 1977 Ron Rivest estimated that factoring a 125-digit product of two primes would take about 40 quadrillion years on the fastest computers of the day. Martin Gardner's August 1977 Scientific American column then challenged readers to break a message protected by a slightly larger, 129-digit number. It was broken in 1994. As Wikipedia's entry on the decrypted message recounts, more than 600 volunteers lent about 1,600 machines to the effort for six months. Two of those machines were fax machines.
Faster hardware accounts for only part of that gap. The rest comes from the condition Rivest stated. The 1994 team used the quadratic sieve, which Carl Pomerance invented in 1981, four years after Rivest's estimate. The asymptotically faster number field sieve came soon after. Rivest was open about his assumption, and the assumption is what failed: that the best method known in 1977 would still be the best method attackers had.
Two of the cryptographers who guided DeepMind's agents know this history first-hand. In 2015 Nadia Heninger and Emmanuel Thomé co-authored Logjam. It showed that breaking 1024-bit Diffie-Hellman mostly comes down to one large precomputation, which can then be reused against every connection that shares the same prime. They put the cost at roughly a $100 million machine running for a year per prime. That is a national budget, not an impossibility. The idea of making factoring cheaper in bulk is older still. Don Coppersmith's 1993 "factorization factory" spread a huge precomputation across many numbers of similar size, although Tanja Lange's later analysis argued that its hardware cost made it impractical. DeepMind's second attack, still unreviewed, belongs to this line of work. It doesn't make any single key easy to break. It makes each additional key cheaper.
What is new here is not the kind of progress but who is producing it. Every key-size recommendation is really a forecast with two parts. The first is growth in computing power, which people track closely. The second is the rate at which new attack ideas appear. That rate is usually treated as slow and steady, limited by how many strong cryptanalysts exist. RSA-129 showed that the second part can matter far more than the first. If agents can produce two new attacks on a heavily studied problem within weeks, that rate is no longer a background constant. It is an AI capability, and it can be measured like one.
This suggests a modest change in how key-size guidance is written. Next to "safe until year X, assuming known algorithms", standards could also say how far year X moves if attack ideas start arriving twice as fast. Rivest wrote his assumption down in 1977, which was the honest thing to do. The lesson isn't to stop writing assumptions down. It's to keep checking whether they still hold.