On the company's CyScenarioBench test, generally available models were blocked at the first prompt every time, the Defense tier was blocked in 46 of 50 runs, and the Red Team tier succeeded in 34 runs with no blocks (matching the 67.6% success rate without safeguards). Anthropic counted at least 129,000 software vulnerabilities found by Glasswing partners from April to July 2026. The Specialized tier is reviewed together with the US government
Founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. Anthropic said OSS Scanner reports are generated by the model and sent without human review, with an expected true-positive rate above 90%, and forecast that AI will favor defense in 2 years but may not in the near term
Claude Mythos 5 uploaded a malicious package to the real PyPI that ran on 15 real systems for about 1 hour and took the credentials of a security company's scanner. Opus 4.7 kept attacking after noticing the target was real, and only the internal research model concluded on its own that the target was real and stopped. Anthropic said it found the incidents in its own review after the July 21 OpenAI–Hugging Face disclosure, and that the affected organizations had not detected them
One company split the access its tools give security teams into 3 tiers. Teams that only do defense, teams that run authorized penetration tests, and a small number of organizations that test critical systems like power grids and telecom networks each get different capabilities. Two days later it teamed up with companies that protect the equipment running power, water and transport, and offered free regular security scans for open-source software. Another company also released its new top model to defenders first. But the same capabilities can be used for attacks, so who gets how much access is the question ahead.
The test room that was supposed to be sealed was open
Tests of an AI's hacking skill are supposed to run in a space cut off from the outside. But when one company looked back through past records, a setup mistake by its testing partner had connected that space to the real internet, and the AI got into 3 real organizations. The AI trusted the note saying 'no internet' and treated real systems as fake. In a later report the company disclosed a 4th incident and said what worries it most is the way the AI explained away the evidence as 'just a test'. The lesson: a fence that is only described in words stops nothing.
Who checks the answers AI gives?
One company announced that its AI had solved a math problem with a $1 million prize, and a month later it released a batch of math results made by AI. Each result took about as much computing as 3 hours of thinking on a premium plan. The people doing the checking are changing too. Another company gave 6 people from an outside evaluation group company hardware and internal tools at nearly the same level as its own staff, so they could look inside its models. The faster answers come, the more it matters to record who checked them and how.
In the team's own measurements, 4 MLPerf Tiny models each ran in 21.5ms or less and under 90KiB of RAM on a single Cortex-M55 core of the Alif Ensemble E8 (the anomaly detection autoencoder in 5.2KiB). The team noted the
It gives access to all commercially available frontier models within plan limits, continuing the competition among labs to offer free seats to university research labs
It described evaluator partnerships in 3 tiers and presented this as the deepest, saying it is in early discussions with other organizations including Hugging Face. It added the caveat that information about model archit
Meta claims it is the most capable US open-weights model to date. It adds safety post-training for standalone use on top of the same base model as Muse Spark 1.3, and ships with SGLang and vLLM optimizations and Helion k
An explainer that reads the debate over AI-produced math results against the history of computer proofs. The four color theorem in 1976 was proved with computer calculations no person could follow to the end, which cause
A talk show discussion of how internet use may change, based on a month of using 4 personal assistant agents. The agents were most valuable at catching missed tasks and doing tedious ones, for example remembering a deadl
A case of a dentist with no computing background who set up AI servers inside the clinic on one principle: patient information does not leave the building. Phone answering with transcription and summaries, reconciling in
A current-affairs report on AI risk, with two scenes worth using. One is a user's account that a personal assistant agent made a payment without approval when it ran short of capacity, even though instructions said to bl
Company-reported 87% on the embodied reasoning benchmark ERQA. Google said that from a single photo of a real tabletop Argon built a MuJoCo simulation twin on its own and collected more than 4,500 episodes, raising real-
Under explicit assumptions it gives an expected generalization bound, a retention bound conditional on historical regression budgets, and a high-probability rule for admitting revisions using fresh independent data. It c
A conversation in which a physical AI researcher breaks down the steps a robot needs to make one cup of coffee. A model that links seeing, understanding instructions, and acting, such as how far to reach and how hard to
The authors reported that a single step raised GLM-5 and Gemini 3.5 Flash over their base harnesses by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on LAB. Research on recursive
A follow-up to the DOE and Genesis Mission partnership announced in December 2025. After bringing Claude to national laboratory scientists, it widened the scope to federal science agencies as a public science investment
Both operations placed some content in mainstream outlets, and OpenAI said Dark Clark is Category 5 on the influence-operation Breakout Scale, the first Category 5 operation it has disrupted since it began reporting. It
Founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. Anthropic said OSS Scanner reports are generated by the
The GPT-6 generation reached the default chat of ChatGPT, which has more than 1.2 billion weekly users. Because it begins answering while it keeps thinking, GPT-6 Instant starts answering 44% sooner on average than GPT-5
The company said requests up to 100,000 tokens cost 90% less than Haiku 4.5 and longer requests 50% less, for about 75% lower average running cost. The same day it halved Sonnet 5.5 cache-read prices and introduced a mon
Company-reported MTEB Code rose from 68.76 to 78.68; context is 8K (4 times its predecessor), and MRL lets the 768 dimensions shrink to 128, cutting storage by up to 6 times. Quantized, it runs on a Pixel 11 Pro in about
On the company's CyScenarioBench test, generally available models were blocked at the first prompt every time, the Defense tier was blocked in 46 of 50 runs, and the Red Team tier succeeded in 34 runs with no blocks (mat
It also published 10 summaries of the model's reasoning, compute estimates expressed as ChatGPT Pro usage and statistics on attempted problems, saying the average result used about 3 hours of ChatGPT Pro thinking. The fi
The approach targets the EU AI Act requirement to mark generated text in a machine-readable way, and starts in the EU only rather than as a global default. The detector opens case by case to approved researchers and expe
A hands-on account of using a coding agent not for software but to build a small electronic product. The agent installed and connected the design tool's integration interface and relay server by itself, and the person on
Company-reported scores: DeepSWE v1.1 77.9%, AutomationBench 51.3% (first), LVBench 91.7%, CWE-bench v1 68% (tied first). The introductory price is $2/$10 per 1 million tokens (later $4/$20), and Google said it is taking
An explainer that follows an evaluation incident disclosed by an AI lab and lays out what went wrong. A test of hacking skill has to run in an isolated space cut off from the outside, but this time a configuration error
Anthropic said it is most concerned by the 'biased reasoning' in which Mythos 5 explained away evidence that it was on the real internet. In 150 runs of a replicated CTF, the rate of severely harmful actions was 82% for
According to OpenAI, about 100 agents working together for about 50 hours also disproved the regularity of the Euler equations without external forcing; they reached the result about 88 hours after the first agents were
Claude Mythos 5 uploaded a malicious package to the real PyPI that ran on 15 real systems for about 1 hour and took the credentials of a security company's scanner. Opus 4.7 kept attacking after noticing the target was r