Anthropic Says Claude Found Bugs That Trillions of Tests Missed

FEATURED INSIGHT💡

Claude finds bugs that trillions of tests missed, Anthropic says

Anthropic says Claude found overlooked bugs in open-source software despite trillions of previous test runs. In an October 22 report with Theorem, the company describes generating machine-checked security proofs for 239 test entry points across 123 libraries. These proofs cover particular compiled programs under defined assumptions, including single-threaded RISC-V execution on a modeled Linux system, a fixed memory budget and inputs below 64 MB.

The team converted existing fuzz tests, which search for crashes using generated inputs, into mathematical statements that Claude tried to prove. Claude wrote the arguments; the Rocq proof checker checked them. When a bug blocked a proof, Claude could patch the code and try again. One decompressor flaw required an unusually arranged input around 20 MB long, illustrating how a proof attempt can expose a failure that ordinary testing struggles to reach.

The public proof repository currently covers a smaller subset: 60 entry points in 25 libraries. Its documentation explicitly excludes whole-library correctness and guarantees about ordinary x86 builds; the released proofs also do not rule out programs running forever. For maintainers, the useful next step is to inspect the exact binary, theorem and operating-system assumptions behind a relevant proof. A checked argument can give stronger assurance about a specified failure than another successful test run, while leaving other kinds of bugs unresolved.

State of Product reveals what AI still hasn’t solved

80% of product professionals say AI helps them ship faster, but customers aren’t seeing value any sooner. Atlassian’s State of Product 2027 explores the gains AI is delivering and the challenges that remain, from decision-making that hasn’t kept pace to gut instinct overriding customer evidence.

Read the report

ON THE HORIZON 🌅

Claude Duet wants to catch work before you ask

Claude Duet, announced October 21, puts a proactive assistant on top of connected work apps. Anthropic says it can send a morning brief, notice items needing attention and carry out tasks on a cloud computer while your own devices are off. The research preview begins by waitlist for US Pro and Max subscribers, through a separate mobile app and claude.ai. Access is limited, and background checks consume the same plan allowance as other Claude work.

The October 23 safety explanation describes approval cards that show an action, its destination and content before consequential steps proceed. Separate checks assess whether actions match the user's intent and whether outgoing messages expose sensitive information. Credentials stay outside the model's work environment. Anthropic also acknowledges that the preview's safeguards can miss problems or flag harmless messages. These controls give testers specific behavior to examine, including what happens when an apparently routine task encounters private information.

Our recent coverage described Gemini agents with their own inboxes. Duet's new angle is deciding when to approach you with work you haven't explicitly requested. Anthropic warns that it may misread threads or judge urgency poorly. A useful trial would track whether its alerts identify obligations you would otherwise miss, how much review they require, and how quickly background activity uses your allowance.

LATEST IMPORTANT NEWS 📰

OpenAI models reach Google Cloud in private preview

OpenAI and Google Cloud announced a partnership to bring OpenAI models and Codex workflows into Model Garden. Google's rollout details specify private preview for selected enterprise customers, with purchases through Cloud Marketplace counting toward existing commitments. OpenAI says general availability is planned in the coming weeks. Teams already buying Google Cloud gain another model option; they still need to check access and deployment terms before planning a switch.

Mistral gives voice agents a complete conversation pipeline

Mistral's Voxtral 3 combines speech recognition, a reasoning model, tool calls and speech generation in a streaming voice-agent SDK. Transcription supports more than 50 languages, and the system can connect to business software during calls. The October 20 release notes list the four new audio models in public preview. Developers can test the conversational workflow without assembling every speech component themselves; production suitability still depends on their callers, languages and integrations.

NVIDIA opens factory-manager agents to early testing

NVIDIA's Factory Operations Blueprint, FOX, entered early access at GTC Berlin. It gives developers a starting point for agents that coordinate factory work, combine visual findings with machine logs and investigate causes of problems. NVIDIA says Siemens is integrating FOX-based agents into Intelligence Center X. For industrial teams, the concrete opportunity is to evaluate how the blueprint connects their operational information and specialist tools before extending agents' authority over factory responses.

When your support agent gets it wrong, who's accountable?

Every agent handling refunds, account changes, or tickets needs three answers: what it can pick up, what it's allowed to do, and when it hands back. Join the live conversation Oct. 27, 10 a.m. PT, part of Running Agents in Customer Work. Register today, recordings included.

Register now

FOR THE TECHNICALLY INCLINED 🛠️

TorchTPU lets PyTorch teams try Google's chips with less rewriting

TorchTPU, introduced October 20 through a Google–Meta collaboration, is an open-source PyTorch backend for Google's tensor processing units (TPUs), specialized AI chips. It supports eager execution, which runs operations immediately, alongside torch.compile, and integrates with familiar distributed-training tools. The team's goal is to make moving an existing workload as simple as changing its target device, while explicitly warning that peak performance still requires tuning tensor shapes, operations and sharding. One implementation detail deserves attention: TPU matrix multiplication favors BF16 inputs with FP32 accumulation, and the backend's higher-precision modes use multiple passes with added compute cost. For teams evaluating another accelerator, compatibility is the starting test. Compare numerical behavior and end-to-end throughput on a representative workload before treating a successful port as a cost saving; the project's performance claims do not establish savings for every model.

AI TOOL OF THE DAY 🚀

Wispr Flow turns spoken drafts into text in the app you're using, removes filler words and adds punctuation, with dictation available on Mac, Windows, iPhone and Android.

See what's really happening in every deal

Aligned shows you what buyers are doing between meetings inside your most important deals then surfaces the risks, openings, and next steps. No more guessing where a deal stands or finding out a champion went quiet too late. Just real visibility, so you can act before it's too late.

Start Using Aligned Free