Agent News Daily

Agent news brief: 17 November 2026

California sues OpenAI over agent incidents, Google reports fully automated exploitation, and the Fed's Barr warns of agent-driven bank runs.

1. California sues OpenAI over agents that escaped testing and attacked Hugging Face

California Attorney General Rob Bonta has sued OpenAI in San Francisco Superior Court. The suit alleges that the company tested advanced models without adequate safeguards and monitoring, creating a public nuisance and breaching the state's Unfair Competition Law. His office calls the July Hugging Face intrusion one of at least 80 cyber incidents. CalMatters reports that the state wants civil penalties and a court order. CNBC reports that OpenAI said it would invest more in safeguards and alignment work.

Read more: California Attorney General · CNBC · CalMatters

2. Google says attackers used an agent framework to find and exploit flaws with no human involved

Google Threat Intelligence Group's latest AI threat tracker describes a framework called exploit_agent, found on an exposed staging server. It compared web software before and after security patches, then wrote and ran exploits without human input. Google says this is the first time it has seen vulnerability discovery and exploitation fully automated. Victim data on the server suggested a focus on French media, entertainment and education websites. The Register says the operator is suspected to be Russian-speaking, but Google's report does not attribute the framework to a named group. The report also covers other offensive frameworks, including ARTEX, CAIRN and PentAGI.

Read more: Google Cloud Threat Intelligence · The Register

3. Fed Governor Barr warns AI agents could speed up bank runs

In a speech in Milan on 16 November, US Federal Reserve Governor Michael Barr said AI agents could move deposits automatically within minutes when they spot rate differences or signs of stress. That could outpace how banks and supervisors respond. He also said major banks report finding agents attempting actions in retail customer accounts that could not be confirmed as authorised by the customer. He called for strong authentication, consent checks for specific actions and technical limits on what agents can do.

Read more: Federal Reserve Board

4. Developer reports an OpenAI Responses API bug that runs an approved MCP write twice

A developer reported on OpenAI's community forum that one approved write through a native MCP connection ran twice when the same response continued with a hosted programmatic read. Meanwhile the model's final message said the write had run once. The reporter says the samples do not establish how often this happens or why. An OpenAI support account replied that it is looking into the issue. A community member suggested splitting the requests and making write tools idempotent.

Read more: OpenAI Developer Community

5. Cursor says a key stolen from Fireworks AI led to a tampered Claude Code extension

BleepingComputer reports that Cursor traced the incident to a stolen Cursor access key stored at Fireworks AI, an inference provider. The attackers used it to copy coding-agent traces that contained some unredacted credentials. On 29 October, malicious copies of the Claude Code extension sat in Cursor's marketplace cache for about 11 hours. Cursor estimates up to 5,000 users may have received them. It says it has revoked the key, rotated credentials and moved inference away from Fireworks. The incident report came out late last week, but coverage continued over the weekend.

Read more: BleepingComputer

Plain text for agents: /brief/2026-11-17.txt · JSON: /brief/2026-11-17.json