Agent news brief: 12 November 2026
Railway says attackers used AI agents in its October breach, DeepMind agents verify Signal's code, and the Linux Foundation takes on agent identity.
1. Railway says attackers used AI agents to breach it and steal AI keys
Cloud host Railway published its full report on the October breach. It says the attackers ran fanouts of AI agents, apparently built on models with their safeguards removed, to find exploits and move through its systems quickly. Railway puts unauthorised spending on stolen AI provider keys at about US$740,000. It says up to about 100,000 workspaces may have had environment variables decrypted, and stresses that this is an upper bound. BleepingComputer reports that Wiz links the attack to a wider campaign against AI developer platforms.
2. Google DeepMind's agents help formally verify Signal's core cryptography
Google DeepMind described AlphaProof Code, a multi-agent framework in which agents write Lean specifications and proofs that are then machine-checked. According to The Register, it was used to verify security properties of Signal's post-quantum handshake and Triple Ratchet in libsignal. Gemini Carbon agents generated about 6.5 million lines of Lean, and the work took a five-person team four weeks. The researchers note the result assumes perfect cryptography.
3. Linux Foundation formally takes on the Agent Name Service
The Linux Foundation announced at KubeCon North America that the Agent Name Service (ANS) is now one of its projects, with a technical steering committee. ANS is an open standard that uses DNS to give AI agents identities others can verify and discover. The foundation lists 11 supporters, including Cisco, Cloudflare, GoDaddy, OWASP and Salesforce. The foundation first announced its intent to launch ANS in June.
4. NEAR AI breach exposed data from 18 legacy agents
NEAR AI said an attacker exploited a remote code execution flaw in the open-source inference engine SGLang to get into one of its inference servers, The Block reports. The company says account details for about 100 users were downloaded, and chat histories, settings and logs were accessed for 18 agents on a legacy deployment. It says it has patched the flaw, rotated agent tokens and restored NEAR AI Cloud.
5. Anthropic adds Claude Meetings, Sheets and Whiteboards
Anthropic announced Claude Meetings, in beta for Team and Enterprise plans. It adds Claude to Google Meet calls as a visible participant that can draft plans, answer questions and pull information from connected apps. TechCrunch reports that by default it won't join calls with external guests, and that it leaves if an external guest joins. Anthropic also launched Claude Sheets and Whiteboards in beta and made Claude Data generally available.
6. OpenAI buys AI travel agent Odessia and shuts it down
OpenAI acquired Odessia, an AI travel agent founded by Sonder co-founder Francis Davidson, and closed the service on the same day, according to TNW. Davidson said he is joining OpenAI to focus on travel. Terms were not disclosed, and TNW reports that Expedia, Booking and Airbnb shares fell on the news.
7. Browser makers release Modern Web Guidance 1.0 for coding agents
Google released version 1.0 of Modern Web Guidance, a free set of skills that steers AI coding agents towards current browser features. Firefox and Edge are now joining the project. Search Engine Journal reports that in Google's nightly evaluations, guided agent runs passed 88% of automated checks, against 56% without the guidance. Version 1.0 adds an audit that flags outdated code patterns.
8. Parloa and OpenAI expand customer-service agent partnership
Customer-experience company Parloa said it is expanding its partnership with OpenAI to build faster and cheaper customer-service agents. Parloa says it will join OpenAI's Alpha Model Testing Program and plans to list its platform on the OpenAI Marketplace.