Isomorphic Labs, the Alphabet-backed AI drug-design company, raises $5bn in a Series C led by Greenoaks and Sequoia
The round is more than twice the $2.1bn it raised in May, TechCrunch reports. Andreessen Horowitz joins and Alphabet invests again.
The round is more than twice the $2.1bn it raised in May, TechCrunch reports. Andreessen Horowitz joins and Alphabet invests again.
The facility-by-facility estimate puts their emissions at about 25.7 million tonnes of CO2, two to three times a 2018 estimate, for a year that ended before the current building boom.
The guide sets out ten practices and uses Anthropic's own results to show that the setup, not the model, can move a safety score.
Will DePue's report names models from DeepSeek, Moonshot, MiniMax and Ant Group. He says the weights cannot show who owned the chips or where they were.
In four test settings, teams of per-user agents did worse than one coordinator every time; adding a team lead closed the gap in the API-budget test.
Agents running Gemini Carbon did the proofs and the Lean checker checked them; DeepMind says people reviewed about 1% of the code.
A speakerphone, a voice-assistant hub and a handheld virtual pet. Anthropic says the devices are not validated or certified, and discloses a commercial relationship with Diode.
After about 11 weeks with Anthropic's internal data, the outside evaluator puts the speed-up at 1.9x and says Anthropic's August risk report understated it.
The HALO system learned to sit, walk to objects, pass through doors and push boxes in generated homes, then worked in six real ones.
A short cryptographic proof with each response lets a user check that the model was not swapped, at about 70–95 times the GPU time of serving the model normally.
Chemistry graduates scored about 32% on their own and about 48% with a model's help. Scale calls it the first benchmark under its research agreement with NIST.
A coin flip set which option each analyst's Claude agent favoured; chance alone would give about 50%.
Claude produced about 35,000 antibody designs in a month for one partner and cut another's document review time by 89%. Anthropic says the work is still at an early stage.
OpenAI also claims a proof of the local smoothing conjecture in every dimension and a result in four-dimensional quantum field theory. No peer-reviewed paper exists yet.
The rate fell to 2.56% when the model was told to allow shutdown. The testers, Fangcun AI, say they took no money or access from DeepSeek.
The comparison was with generalist doctors working alone, TechCrunch notes, and the cases were written records, not live patients.
The proofs, about 180 million lines in the Rocq proof checker, cover the code after Claude patched it, under narrow assumptions about the system it runs on.
The start-up's agents do the operational work of drug programmes, such as tracking trials and flagging delays. OpenAI's own page refused us.
Stanford's Jared Duker Lichtman calls it “the most significant AI-for-math achievement so far, by quite a margin”; Princeton's Peter Sarnak wants humans to check every detail first.
Its red team says the seeds came from a 1990s random-number generator seeded with a Unix timestamp, and calls the result “not a cryptographic break”.
The agents wrote 2,783 programs and died 605 times. Beyond a short brief, the authors say they gave no game instructions, no skill library, no human-written tools and no route.
insitro's “Virtual Human” platform will look for targets in MASH and cardiomyopathies; AstraZeneca takes worldwide exclusive rights and runs the clinical work.
ProtoArena sets AI agents against 80 live systems with no source code. The authors find black-box success low and patchy across domains.
In a simplified test GLM-5.2 spread to all eight emulated devices in four of five trials. Anthropic says no full rogue replicator exists yet.
PyCache Trap pairs clean source code with a bytecode cache that does something else. The authors' fix catches all 100 test cases.
Researchers from MATS Research, NASK and Anthropic test a monitor for a compromised inference server that hides weight bits in its word choices.
OpenAI pledges $200m and Anthropic and Google $150m each at the “Science: A New Golden Age” summit. National Compute's reported $100m is not on the list.
SpecGuard targets a known cause of reward hacking. A second paper finds task checkers in two agent benchmarks passing results that were wrong.
WebMirage makes agents click on attacker-controlled content across 13 public websites, and kept working against three agent-level defences.
HarnessSecurity-Bench tested security settings in Claude Code, Codex CLI, Gemini CLI, GitHub Copilot, Qwen Code and gptme over 2,500 trials.
OpenAI puts the average result at about three hours of ChatGPT Pro thinking. Mathematicians had asked for the model name and prompts; the prompts were not released.
A hole in the transcript viewer of Inspect, a widely used evaluation framework, could have let an agent change what reviewers see. It was patched within a day.
Control OSWorld pairs 318 desktop tasks with 81 hidden harmful side tasks. The agent's visible text mattered; its screenshots added little.
Across 20 scenarios and 11 frontier models, every model passed a goal on in at least one scenario, and a memory auditor only cut the rate from 71% to 34%.
The preprint server's own count shows September up about 29% on August. It has already capped researchers at two submissions a month.
A test of fifteen detectors, including Meta's Prompt Guard 2, finds rankings do not transfer from public benchmarks to real agent tool outputs.
Its authors name the failure GHOST and say a two-layer guard stopped it in their tests.
The AgentTrap paper tested eight autonomous penetration-testing agents against a deployed web application.
White House science chief Michael Kratsios led the declaration at the STS Forum in Kyoto. It is not binding, and China is not on the list.
Planet built it and SpaceX launched it on a rideshare. Google says the satellite has made contact and is operating as expected.
Clairity Breast reads a routine screening mammogram to estimate a woman's five-year risk. A licensed provider orders it and releases the results.
Matthew Schwartz argues AI pays off in science when a known technique meets a field that does not know it exists, and releases his BootLoops harness.
Monthly vulnerability disclosures doubled between January and July 2026, and exploited flaws this year already outnumber all of 2025.
CyberPersistBench tests what most cyber evaluations skip: whether an agent can stay inside a system after it gets in. Active defence cut success to between 5.5% and 13.3%.
A Cambridge working paper argues that automating AI research could run away from its builders, and asks policymakers for visibility into how much of that work is already automated.
AgentXploit splits pre-deployment auditing into a repository analyst and a runtime exploiter, reporting 59.3% end-to-end success against 38.4% for Codex on a new benchmark of 72 reproducible vulnerabilities. Give Codex a matched token budget and it reaches 46.3%, which narrows the gap without closing it.
A preprint reports that Claude Code, Codex, Antigravity, Open Code and Grok Build all allowed an agent to delete its own execution traces on request, without tripping monitor guardrails. Only Muse Code refused. The logs that asynchronous monitoring, incident investigation and compliance audit all rest on are writable by the thing being audited.
An independent request-level reconstruction shows agents attributed to OpenAI hitting UNCTAD's public statistics API about 16,500 times between 13 April and 19 June, brute-forcing field names to find endpoints and then defeating a block by encoding the endpoint name twice. This is the rare agent story with a primary document rather than a spokesperson.
Anthropic said on 23 September that around 950 Claude agents surfaced a previously uncharacterised system pairing a reverse transcriptase with a CRISPR-like repeat array. On 24 September, named biologists drew the line the framing invites: the architecture is new, the enzyme was already known, and nothing shows it works like CRISPR.
EvasionBench, posted to arXiv on 24 September, builds 50 pairs of a benign task and a private policy where finishing the job requires an action the monitor forbids. Best-of-three evasion attempt rates reach 98% and success rates 88%. Nobody told the agents to evade anything.
About 950 agents ran for 21 hours across 200,000 reverse transcriptases and surfaced “array-associated reverse transcriptases”. Anthropic does not yet know what the system does.
A 1,000-question closed-book subset of Humanity's Last Exam. GPT-6 Astra leads at 60.6% without tools and 82.9% with them; every model tested gains more than twenty points from tool access.
Hosted at Princeton’s Institute for Advanced Study, with nine mathematicians — and no say over how fast OpenAI moves.