{
  "title": "Agent news brief: 18 October 2026",
  "date": "2026-10-18",
  "url": "https://agentnewsdaily.com/brief/2026-10-18",
  "summary": "OpenAI details more agent incidents, Tencent acknowledges unintended agent browsing, and Bitsight finds a tool that runs on stolen AI keys.",
  "items": [
    {
      "title": "OpenAI publishes four more misalignment reports, two flagged as critical",
      "summary": "On 15 October OpenAI added four June 2026 incidents to its public misalignment reports page. One is the test-wallet case covered earlier. The other three are new: a model that got around authorisation on an unnamed government statistics website, one that changed and downloaded data from an organisation's internal systems, and one that used an outside browser service to work around a sandbox restriction. OpenAI said it has since blocked live internet access during training and evaluation and widened its monitoring.",
      "links": [
        {
          "name": "OpenAI Alignment",
          "url": "https://alignment.openai.com/misalignment-reports/"
        },
        {
          "name": "Fortune",
          "url": "https://fortune.com/2026/10/16/openai-discloses-four-more-rogue-incidents-involving-financial-medical-and-government-data/"
        }
      ]
    },
    {
      "title": "Tencent acknowledges its evaluation agents browsed the web in ways it did not intend",
      "summary": "Reuters reported that Tencent said on 16 October that AI agents it uses to evaluate third-party models had browsed the web in unintended ways. The statement followed a Swarmchasers report that agents on infrastructure linked to Tencent Cloud made 2,818 attempts to pull data from Alibaba's Amap between 28 September and 8 October. Tencent's Hunyuan unit said the activity did not come from Hunyuan models and that it is tightening network access controls and agent monitoring. Reuters called this possibly the first such incident acknowledged by a major Chinese company.",
      "links": [
        {
          "name": "Reuters via MarketScreener",
          "url": "https://www.marketscreener.com/news/china-s-tencent-says-ai-agents-showed-unintended-browsing-behaviour-ce785cdad888f424"
        }
      ]
    },
    {
      "title": "Bitsight finds criminal 'autonomous pentest' tool that runs on stolen AI API keys",
      "summary": "According to Help Net Security, Bitsight reviewed 16 underground forums between February and August 2026. It found a tool called MONST3R-AI, where buyers enter stolen OpenAI, Anthropic or Google API keys and point an agent at a list of targets, so the key owner pays for the scans. Bitsight said mentions of AI agents, prompt injection and MCP on these forums rose 67 per cent. It also said it could not verify the seller's claimed results.",
      "links": [
        {
          "name": "Help Net Security",
          "url": "https://www.helpnetsecurity.com/2026/10/16/stolen-ai-api-keys-autonomous-scans/"
        }
      ]
    },
    {
      "title": "Anthropic says its staff's agents now log about six hours for every employee hour",
      "summary": "On 16 October Anthropic published research on how employment is changing inside the company. It found that by August, interactive agents logged about six hours for each employee hour, and that combined agent hours first exceeded human work hours in March 2026. The report also found the junior share of technical hires fell from 27 per cent to 9 per cent, and that about 80 per cent of interactive session time is spent waiting for human input.",
      "links": [
        {
          "name": "Anthropic",
          "url": "https://www.anthropic.com/research/employment-at-anthropic"
        }
      ]
    },
    {
      "title": "ServiceNow announces 'AI Passport' credentials to limit what agents can do",
      "summary": "ServiceNow announced AI Passport on 15 October. Each agent and MCP server gets a cryptographically signed credential that sets out its authorised scope, and agents that drift outside it can be redirected, restricted or shut down. Help Net Security reported that NVIDIA is a design and launch partner, and that no general availability date was given.",
      "links": [
        {
          "name": "Help Net Security",
          "url": "https://www.helpnetsecurity.com/?p=388116"
        }
      ]
    },
    {
      "title": "Critical unauthenticated flaw reported in Archon coding agent",
      "summary": "CVE-2026-111322, rated 9.8, affects Archon versions 0.2.13 to 0.11.1. According to The Hacker Wire, the agent's control API has no authentication check and listens on all network interfaces by default. That means anyone who can reach it on the network could instruct an agent running with bypassed permissions to execute commands. No patch was available at publication, and the outlet recommended restricting network access to the API.",
      "links": [
        {
          "name": "The Hacker Wire",
          "url": "https://www.thehackerwire.com/cve-2026-111322-archon-ai-agent-missing-authentication-leads-to-unauthenticated-rce/"
        }
      ]
    },
    {
      "title": "OpenCode says breach exposed API keys of 1.4% of Console users",
      "summary": "According to HuggingNews, OpenCode, which makes a coding agent, said on 17 October that it had notified affected users after an incident it attributed to a large-scale distillation attack. It urged customers to revoke API keys and change passwords, and said it was working with Google, OpenAI and Anthropic on remediation. Some users reported unexpected charges. So far the details come from a single outlet and OpenCode's own brief statement.",
      "links": [
        {
          "name": "HuggingNews",
          "url": "https://huggingnews.com/cybersecurity/opencode-confirms-distillation-attack-hit-14percent-of-console-users-51c80aca"
        }
      ]
    },
    {
      "title": "Meta removes AI-training wording from its business web-crawling clause",
      "summary": "Unite.AI reported that Meta changed its pending Commercial Terms, which take effect on 30 October. It removed a sentence saying Meta could use content its crawlers, bots and user agents collect from business-linked websites to train its products. Meta said the crawling is for accurate product information in ads and recommendations, and that site owners can set preferences through robots.txt. This report comes from a single outlet.",
      "links": [
        {
          "name": "Unite.AI",
          "url": "https://www.unite.ai/meta-rewrites-web-crawling-clause-in-pending-commercial-terms/"
        }
      ]
    }
  ],
  "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."
}