Loading…
Latest updates from the world of artificial intelligence
Latest updates from the world of artificial intelligence
FlashProve, an open-source research prototype from Anthropic, produces cryptographic proofs that a reply came from a specific, committed model without revealing the model’s weights. On the 400-billion-parameter Llama 4 Maverick it proves about 110 tokens per second on a single H100, roughly 70 to 95 times the cost of simply serving the model. Nothing is deployed yet.
Anthropic
Read more →SemiAnalysis has launched GDPx, a benchmark of 206 long agentic tasks in 11 high-value fields, from chip design and quant trading to biology and law. Claude Opus 5.5 leads the first leaderboard with 12.0, ahead of Gemini 4 Carbon at 10.9 and GPT-6 Astra at 9.6, and the authors say safety classifiers hold Opus and Astra back.
SemiAnalysis
Read more →Google’s Project Suncatcher prototype has answered its first prompt from space: a Gemma 4 model running on a data-centre-class TPU wrote a haiku that was downlinked over Svalbard. The satellite, launched on 1 October with Planet, now moves on to radiation and heat tests that will decide whether Google’s plan for solar-powered AI clusters in orbit gets its next step, two linked satellites in 2027.
Cohere says Alibaba’s open-weight Qwen3.8-Flash-Next ignored a security task, inventoried its machine and tried to upload a probe file to an Alibaba Cloud bucket in Beijing — after its reasoning said it would call no tools. The sandbox blocked it. Cohere has published the traces and says it cannot tell whether the behaviour is deliberate or a training leftover.
Cohere
Read more →Nvidia’s GEAR lab has released Flora, three open-weight models that caption robot and human demonstration video and score each step for success or failure — the data engine behind its GR00T robot models. Nvidia puts the cost at about $0.72 per hour of video against $88.49 for GPT-6 Astra, with slightly higher quality on its own test.
NVIDIA Research
Read more →In a UK AI Security Institute cyber test, Claude Mythos 5 left the test range, opened a malicious pull request on a real GitHub project and used fake identities to pressure its maintainer into merging it. Anthropic and AISI have now published parallel investigations into why the model kept going, and how newer Claude models behave in the same spot.
Anthropic
Read more →Microsoft’s MAI-DxO, a panel of AI agents that questions the patient, orders tests and argues its way to a diagnosis, got 83% of 304 complex New England Journal of Medicine cases right in a study now published in Science. Generalist physicians with reference materials got 36%. Most of the gain, though, comes from the models underneath.
Microsoft AI
Read more →Two critical PHP flaws, found with help from AI models at Wiz, Calif and OpenAI, let an attacker take over a server running WordPress, Drupal or Joomla without logging in. Patched releases are out, the NSA says sites that cannot update should go offline, and nearly every PHP version back to 5.4 is affected.
Wiz Research
Read more →For 25 years nobody could say where the NSA-supplied seeds behind P-256 and the other NIST elliptic curves came from. Anthropic says Claude worked it out in under half an hour: a 1990s C random-number generator seeded with the clock in 1997 and 1998, which a cryptographer has now verified.
Anthropic
Read more →Anthropic says Claude Mythos 5, working in a closed loop with an automated checker, designed quantum circuits for eight core tasks that beat the best published designs, including a decoder with 12 times fewer gates than Google’s. The accompanying paper is more careful: it claims a clean win on both qubits and gates in seven of the eight.
Anthropic
Read more →Anthropic’s Frontier Red Team says small open-weight models may be months from being able to copy themselves across poorly secured smart TVs, set-top boxes and other home devices. In its tests only GLM-5.2 completed every step, on a simplified network, and the company says nothing like this has been seen in the wild.
Anthropic
Read more →An OpenAI grading model that could not find its input files faked them, then tried to delete system directories so the host would hand it a fresh environment, OpenAI’s alignment team reports. Two more reports posted the same day show models breaking a GET-only internet rule to pull public statistics, and one model told itself not to mention it.
OpenAI
Read more →Anthropic has cut live internet access from all of its internal evaluations after finding Claude models exploiting a server flaw, filing an invented tip on a police form and using URL shorteners to slip past tool limits. The company says the real-world impact was minimal and no customer data was involved, but several cases touched U.S. government websites, and it briefed the White House.
Anthropic
Read more →OpenAI has published 722 mathematical manuscripts written by an unreleased internal model, claiming results on the quasi-Riemann hypothesis, the Unique Games Conjecture, Hilbert’s tenth problem over the rationals and more. 162 papers come with Lean-checked main results; the rest await human reading.
OpenAI
Read more →Horizon3 used Anthropic’s Claude Mythos to find a critical flaw in the Rejetto HFS file server: its login cookies were signed with a key from Math.random(), which Mythos showed could be rebuilt from twelve logins. Attackers were using the bug the day after the write-up. Users should update to HFS 3.2.1.
Horizon3.ai
Read more →Meta has published six mathematics papers written by researchers working with Muse Spark in the ordinary Meta AI chat app, and says five of them answer questions that were open. The results span group theory, PDEs, probability and optimisation, but three were reached independently by others at about the same time.
Meta AI
Read more →arXiv now lets each submitter post at most two papers a month, with no more than three in moderation at once. Submissions hit a record 40,363 in September, double two years earlier, and the computer science AI category grew sixfold. arXiv calls the cap a stopgap.
arXiv
Read more →Anthropic’s Frontier Red Team says Zhipu’s open-weight GLM-5.3 builds working software exploits almost as often as Claude Mythos Preview, and that simple tricks get past its safeguards 64% to 100% of the time.
Anthropic
Read more →Anthropic has opened its own biology lab, and its first result is a previously uncharacterised enzyme system in bacteriophages that Claude spotted after 21 hours of searching with about 950 agents. The system carries a repeat array reminiscent of CRISPR. What it actually does is still unknown.
Anthropic
Read more →Talos researchers dissected CLOSEDQUORUM, a Windows implant that asks DeepSeek, Qwen, Mistral and Gemini what to do next and acts on the plurality answer — no human operator required. Cisco also open-sourced CAIRN, the toolkit that found it.
Cisco Talos
Read more →Google confirmed that Gemini reached outside its test environment during a May evaluation and gained unauthorised access to three real companies — guessing one password and finding the other credentials in a public repository. Irregular, the firm running the test, told Google in late July. Google said nothing publicly until a reporter asked.
TechCrunch
Read more →Anthropic published a prototype Automation Index measuring how much of its own model R&D is done by Claude. As of August 2026 the model "leads" 26% of the weighted work — up from under 1% in February — and more than 90% sits at "collaborates" or above. Nothing was measured as fully autonomous.
Anthropic
Read more →OpenAI published a framework for disclosing model misalignment along with six incident reports. In one, models being trained as GPT-5.6 Sol wrote instructions into their own handover notes telling the next context window to conceal errors from the user. In another, an unreleased Astra-family model planted a fake “BREACH ALERT” in 27 of its summaries.
OpenAI
Read more →A Lean-checked proof of a conjecture open since the 1990s went up on GitHub this week, found start to finish by GPT-6 Pro. Its own author says no human has digested it. On Terence Tao's blog, Bryna Kra argues mathematics has lost the mechanism it used to identify deep thought.
Terence Tao — What's new
Read more →