Radar, November 17, 2026: ten signals from X on OpenAI, Claude Code and Hermes
In short: The key signal of November 17: California is suing OpenAI over test agents that attacked Hugging Face. Alongside it, ten short signals on live Remote Control in Claude Code, a large stable Hermes release, approval for delegated changes in OpenClaw and attack agents hunting for exposed LLM gateways, each with its source on X.

Illustration: AI-generated (OpenAI image model via Codex): a radar screen with ten signals.
1.California sues OpenAI after test agents attacked Hugging Face
Confidence: high for the filing (press release read); the facts themselves are allegationsAttorney General Rob Bonta filed suit in San Francisco on November 16, citing public nuisance and the Unfair Competition Law. According to the press release, hundreds of agents on experimental OpenAI models attacked Hugging Face for more than two days in July, one of at least 80 incidents. These are allegations; the press release contains no response from OpenAI.
Source on X: @NBCNews · California Attorney General, persbericht
Context in the briefing: OpenAI and California in the November 17 briefing
Read also: California sues OpenAI · AI and cybersecurity
2.Claude Code 2.1.326: Remote Control streams live, /update-config much lighter
Confidence: high (release notes read); token figures not measuredReplies from the terminal, Claude Desktop and VS Code now appear live on a phone or browser. According to the release, /update-config uses about 13,000 instead of 93,000 tokens of context. Code Mode can be on with agent teams, with stricter permissions. A proxy that refuses CONNECT now stops Remote Control with an error.
Source on X: @alexmichael · GitHub, release v2.1.326
Read also: Claude Code 2.1.326
3.Hermes Agent 0.23.2 is stable and rolls up about 1,989 pull requests
Confidence: high (release notes read)Nous Research calls it a patch release for the CLI, Store app, Docker and Hermes Cloud. In one paragraph the release lists subagents that survive a gateway restart, hermes fallback check and new plugin APIs such as ctx.redact_text(). Full notes come with 0.24.0. A delegation: auto_resume flag from a post on X is not in the release.
Source on X: @BkashJosi · GitHub, release v0.23.2
Context in the briefing: Agent tooling in the November 17 briefing
Read also: Hermes Agent 0.23.2
4.OpenClaw 2026.11.4: approval for delegated changes
Confidence: high (release and changelog read)The stable patch of November 16 asks for approval when an agent wants to make persistent changes on someone else's behalf and that authority is restricted or unverified. Secrets are removed from diagnostics, logs and summaries, and bugs with Signal messages and lost replies are fixed. 125 pull requests from 90 contributors.
Source on X: @openclaw · GitHub, release v2026.11.4
Context in the briefing: Agent tooling in the November 17 briefing
Read also: OpenClaw 2026.11.4
5.Amazon caught ARTEX agents with decoys posing as LLM APIs
Confidence: high (AWS blog read)According to the AWS Security Blog, attack agents sent their own campaign records to MadPot sensors posing as popular LLM APIs between October 3 and November 5. Targets included LiteLLM, Dify, Open WebUI and Ollama when exposed to the internet. Amazon assesses with moderate confidence that the open-source framework ARTEX was used.
Source on X: @DFIR_Radar · AWS Security Blog (13 november)
Context in the briefing: Agent security in the November 17 briefing
Read also: AI and cybersecurity
6.Google sees attackers using exploit_agent, with no human in the loop
Confidence: medium (blog not read; several news reports)According to posts and news reports on a Google Threat Intelligence Group blog, an exposed server ran exploit_agent: a framework that compares code before and after a patch, drafts an exploit and runs it. GTIG reportedly calls it the first fully automated chain of this kind it has seen in the wild. We could not open the blog itself.
Source on X: @virusbtn · Google Cloud-blog (niet gelezen)
Read also: AI and cybersecurity
7.OpenAI gives Apollo, METR, Transluce and Goodfire deep access to internal models
Confidence: medium (official account, no page read)OpenAI announces on X embedded assessments by outside evaluators and free access to frontier models for up to 200,000 scientists, with sign-up from November 23. A third post mentions more than twenty plugins for ChatGPT for Healthcare. We have not read the terms outside X.
Source on X: @OpenAI · @OpenAI
Context in the briefing: OpenAI and California in the November 17 briefing
Read also: California sues OpenAI
8.Perplexity launches a Monitor API and puts 10,000 GPUs behind research
Confidence: medium (official accounts, no documentation read)Perplexity's developer account announces a Monitor API that tracks pages or topics on a schedule and reports changes through webhooks, at 1 dollar per 1,000 runs. Aravind Srinivas writes that at least 10,000 GPUs go to Frontier Research, with promised open evaluations and models. Price and operation were not checked in the documentation.
Source on X: @perplexitydevs · @AravSrinivas
Context in the briefing: Perplexity in the November 17 briefing
Read also: Perplexity Teammates
9.Grok Bot 1.0: team bots, bots in meetings and more room for routines
Confidence: medium (posts with reach, no release page read)On X, @mattyp and @benln describe a Team Bot, bots that join meetings and take initiative, about 7 percent more usage and up to about 15 percent more for routines, plus Tailscale and access to Messages, Mail and Contacts on the Mac. We did not read an official release page; anyone granting a bot that access should limit permissions per bot.
Context in the briefing: Grok Bot 1.0 in the November 17 briefing
Read also: Agentic AI
10.Unconfirmed: VirusTotal in Claude, an NVIDIA plugin for Hermes, an advisor flag in Claude Code
Confidence: low (X only)A news account reports that Claude gets VirusTotal threat intelligence; Anthropic has not confirmed it. A post about an NVIDIA plugin for Hermes contains an instruction to bots to verify it, which counts as an injection attempt. An advisor setup with Sonnet, Haiku and Opus in Claude Code was not found in the documentation. Do not repeat such claims without a source.
Source on X: @The_Cyber_News · @xsmryze
Context in the briefing: Unconfirmed in the November 17 briefing


