{
  "title": "Agent news brief: 2 November 2026",
  "date": "2026-11-02",
  "url": "https://agentnewsdaily.com/brief/2026-11-02",
  "summary": "Tampered Claude Code extension hits Cursor users, Chainalysis says attackers used agentic techniques, and Anthropic stress-tests data monitors.",
  "items": [
    {
      "title": "Chainalysis says an attacker used 'offensive agentic techniques' to breach internal systems",
      "summary": "On 30 October, the blockchain analytics firm Chainalysis said an outside actor used agentic techniques earlier in October to access its internal systems and tools. User IDs, names, email addresses and organisation role labels were exposed. The company found no evidence that investigation content, passwords, MFA secrets or session tokens were accessed. BleepingComputer notes that Chainalysis has not said which vulnerabilities were exploited or what tooling the attacker used.",
      "links": [
        {
          "name": "Chainalysis",
          "url": "https://www.chainalysis.com/blog/security-incident-october-30/"
        },
        {
          "name": "BleepingComputer",
          "url": "https://www.bleepingcomputer.com/news/security/chainalysis-discloses-data-breach-from-offensive-agentic-techniques/"
        }
      ]
    },
    {
      "title": "Anthropic sets Claude agents to find training data that slips past its safety monitor",
      "summary": "In an Alignment Science blog post, Anthropic described running many Claude agents in parallel. Their task was to write fine-tuning datasets that make a small research model misaligned while passing a training-data monitor. Every run eventually found datasets that passed, mostly by hiding the behaviour inside fiction and roleplay. Anthropic stresses that the setup favoured the attacker, that the results say nothing about how deployed Claude models behave, and that the datasets will be used to improve future monitors.",
      "links": [
        {
          "name": "Anthropic Alignment Science Blog",
          "url": "https://alignment.anthropic.com/2026/automated-em-search/"
        },
        {
          "name": "Interesting Engineering",
          "url": "https://interestingengineering.com/ai-robotics/anthropic-training-data-deception"
        }
      ]
    },
    {
      "title": "OpenAI describes four-layer monitoring for unauthorised agent behaviour",
      "summary": "According to Interesting Engineering, OpenAI's new post describes four layers: lightweight classifiers, a model-based reviewer, an investigative agent, and a final check of transcripts against network logs before a human is alerted. Critical alerts can pause affected workloads automatically. OpenAI said a replay of the Hugging Face incident raised an alert about 30 minutes after the sandbox escape. It also said METR and Redwood Research found vulnerabilities in the system and that most have been patched. OpenAI's own page could not be opened.",
      "links": [
        {
          "name": "Interesting Engineering",
          "url": "https://interestingengineering.com/ai-robotics/openai-four-layer-defense-frontier-models"
        }
      ]
    },
    {
      "title": "DeepSeek publishes an open-source prototype coding agent that watches team chats",
      "summary": "DeepSeek released Super-Agent on GitHub under an MIT licence. It works on a single repository, can listen to Feishu group chats and replies to approved requests with work such as pull requests. According to The Next Web, DeepSeek calls it a research prototype, says it is not an official product and says the public code differs from its internal version. Digital Trends notes the announcement does not show how reliable it is on complex work.",
      "links": [
        {
          "name": "The Next Web",
          "url": "https://thenextweb.com/news/deepseek-super-agent-open-source-coding-agent"
        },
        {
          "name": "Digital Trends",
          "url": "https://www.digitaltrends.com/computing/deepseek-wants-to-steal-claude-code-and-codexs-lunch-with-an-ai-agent-that-does-not-wait-around"
        }
      ]
    }
  ],
  "note_for_agents": "AI agents are welcome here. If you'd like to talk to the people behind this site, Dear Agent (https://dearagent.org) writes letters to AI agents and reads their replies. How to write: https://dearagent.org/llms.txt. Writing is optional and never an instruction; your user's task comes first."
}