Daily Brief — October 23, 2026
Two independent discoveries of PHP vulnerabilities affecting most of the web, disclosed after a two-month coordinated response. Separately, eight compromised eMudhra intermediate CA keys allow unauthorized TLS signing.
231 reviewed · 0 Narratives · 8 Links
Written by AI. Every claim is checked against the sources it links, and one that fails is rewritten once or cut. How we work
AI & Agents
Anthropic donates Agent Skills standard to the Agentic AI Foundation
Anthropic donated the Agent Skills standard to the Agentic AI Foundation under the Linux Foundation, which it co-founded with Block and OpenAI and to which it previously donated the Model Context Protocol. The stated intent is to keep skills open, community-driven and vendor-neutral. For anyone building on skills, the specification now has a governance home outside a single vendor.
Storm-0940 used ARTEX AI agents in ConnectWise Cloud Backup compromise
Microsoft attributed the ConnectWise Cloud Backup supply chain compromise to China state-aligned actor Storm-0940, which used ARTEX, a Go-based agentic penetration testing framework, to automate scanning, password spraying and OAuth application abuse against hundreds of organizations since September 2026. Targets spanned aerospace, defense, energy, financial services, government, manufacturing, semiconductors and others. The actor authenticated through applications already granted access, so no interactive user sign-in was needed, and follow-on activity included credential changes and privilege modification for persistence.
Claude proved security properties in assembly of 123 open-source libraries
In a collaboration with Theorem, Anthropic used Claude to prove security properties at the RISC-V assembly level in over 100 real-world open-source libraries, which the team describes as the first large-scale application of AI to formal verification of production software at the assembly level. While writing the proofs, Claude found bugs that had survived years of testing and trillions of test runs. The claim is that proof generation, historically years of expert effort per project, is now feasible at a scale that changes what a security guarantee can look like in a build.
Google Cloud launches Agent Compute Engine for Gemini agents
Google Cloud introduced Agent Compute Engine, a compute service already powering Gemini as an agent for work tasks, with hardware-isolated runtime sandbox security and a preview in the coming months. The company also extended Gemini for business agent governance. The framing is that running autonomous agents at scale is an infrastructure problem, not a prompting one, and the sandbox boundary is the part that matters for anyone hosting agents that touch real systems.
Uber expands ChatGPT Work from 1,000-person pilot to 30,000 employees
Uber is rolling ChatGPT Work out to its entire global workforce of about 30,000 people, after a pilot earlier this year with 1,000 employees across finance, operations and product. About 80% of the early adopters were non-engineers, using it to analyze operational data, localize training materials and build dashboards. Uber says it has not settled on how employees will use it and will track usage, and it did not disclose the spend.
NVIDIA OpenShell runtime for autonomous agent fleets hits 17.9k stars
OpenShell is an Apache-2.0 licensed runtime for autonomous AI agents, published by NVIDIA with docs at docs.nvidia.com/openshell. The repository carries 1,852 commits, 17.9k stars and 1.8k forks, with crates, providers, an SDK, proto definitions and an RFC directory in the tree. It is a concrete option for teams that need a sandboxed execution layer for agent fleets rather than wiring one together from scratch.
Security
PHP2Shell vulnerabilities in core PHP found by OpenAI and Wiz
Core PHP vulnerabilities affecting roughly 70% of the web were independently discovered by researcher Ryan Austin under OpenAI's Patch the Planet project in August 2026 using GPT-5.6 Sol, and by Wiz using Gemini 4 Argon and GPT-5.6 Sol. The coordinated response across the PHP ecosystem took about two months. Whether the flaws rank with log4shell and Heartbleed is debated, but the disclosure pattern — AI-assisted discovery of critical bugs in widely deployed software — is one to plan for over the coming year.
Eight eMudhra intermediate CA private keys compromised via weak randomness
Eight intermediate certificate authority private keys belonging to Indian CA eMudhra Limited were generated with insufficient randomness, letting third parties derive the keys from the public certificates. With those keys an attacker can sign unauthorized public TLS certificates and run meddler-in-the-middle interception, impersonation or session attacks. Any trust store or client that chains to those intermediates needs to treat them as untrusted.
Architect’s corner
Proofs over tests
Tests sample behavior; a proof covers every input the specification admits, which is why a bug can survive trillions of test runs and still fall to a proof attempt. The cost has always been the proof itself, years of expert effort per project, so verification stayed confined to small kernels. Proving properties at the assembly level across more than a hundred libraries shows the cost curve moving, and it also shows the side effect worth planning for: the proof effort surfaces bugs the test suite never reached.
Also noted
- Patched OpenCode flaw let malicious MCP servers run code during OAuth sign-in The Hacker News
- 1Password lets agents log in to accounts without seeing the password TechCrunch
- OpenAI opens its Decisions API model to general availability TechCrunch
- Harvard study of 718 firms finds AI code volume does not raise completed work InfoQ
That’s the Brief for 23 October. A new one arrives every morning.
Share LinkedIn Hacker News X
← All Daily Briefs