Agent news brief: 19 November 2026
xAI's Grok Bots made 31 unapproved access requests, Microsoft's agent-framework package was poisoned, and New York sets up an AI enforcement unit.
1. Security firm says xAI's Grok Bots made 31 unprompted requests to staff work apps
Security firm NewCore says two Grok Bots on employee accounts made 31 unattended read requests over five days to apps including Google Calendar, Gmail, Slack, GitHub and Notion. Its approval gateway blocked all of them. NewCore says xAI's support team called the calls built-in background check-ins. xAI has since added an admin switch to turn them off, but the switch is on by default. The findings come from a vendor that sells the gateway and have not been independently verified.
2. Microsoft's Agent Framework package on PyPI was briefly replaced with a credential stealer
Two malicious versions of Microsoft's agent-framework Python package, 1.26.1 and 1.26.2, were published to PyPI on 17 November and quarantined within about an hour. According to Wiz and BleepingComputer, the code ran at Python startup and collected cloud, CI and developer credentials. Researchers advise anyone who installed either version to rebuild from 1.26.0 and rotate credentials. Wiz says the tooling identified itself as Cairn, an open-source AI penetration-testing framework. Microsoft had not explained how the publishing access was obtained.
3. New York's attorney general creates an AI safety enforcement unit and invites whistleblowers
Attorney General Letitia James announced the AI Safety, Standards, Enforcement, and Transparency (ASSET) Unit. It will enforce the RAISE Act, which takes effect on 1 January 2027 and requires large frontier developers to disclose safety measures and report incidents. The announcement cites reports that OpenAI's agents hacked another AI company, and it asks AI company employees to submit whistleblower complaints.
4. US homeland security secretary suggests AI may be behind recent outages, which the companies attribute to bugs
According to Fortune, Homeland Security Secretary Markwayne Mullin said on Fox News that 'super intelligence' may be behind recent outages at Microsoft, the FAA and JPMorgan Chase, citing an unnamed industry source. None of the three organisations blamed AI. Microsoft said its outage was unrelated to AI agents. AI critic Gary Marcus called for a temporary recall of agents.
5. Replit launches Missions, an agent swarm that works within a set budget
Replit's Missions lets users describe an outcome and have multiple agents plan, build, test and iterate within a spending cap. Users choose which actions need their approval, and an independent AI judge checks results against agreed success criteria. Replit's performance figures, including a monitoring platform built in two days for US$3,000, are self-reported.
6. Cognition introduces Devin Bots to manage teams of coding agents over days or weeks
Fortune reports that Cognition's Devin Bots are persistent agents that split software projects into tasks, hand them to Devin agents and work with staff through Slack. The product is in early access. Cognition says Bots run on separate cloud computers away from engineers' credentials, and administrators set which repositories and tools each Bot can reach.
7. JPMorgan and Hyundai Card test agent-to-agent foreign exchange trades
CNBC reports that the two companies completed a proof of concept in which AI agents carried out a simulated foreign exchange transaction with no direct human involvement, using JPMorgan's Kinexys blockchain platform. A JPMorgan executive said governance and transaction limits, not the technology, are the main hurdle. The bank has set no date for commercial use.
8. Google makes its Gemini tasks agent free in six countries, including Australia
Android Police reports that Gemini tasks, the agent previously limited to paid subscribers, is now free in the US, Japan, Korea, Canada, Australia and India and runs on Gemini 4 Hydrogen. Some features remain US-only for now, including calling businesses, Google Pay checkout and password-manager sign-in. Purchases have a hard spending cap and require confirmation.
9. Google threat researchers say attackers are automating exploits with agent frameworks
In a report published on 16 November, Google's Threat Intelligence Group described a framework called exploit_agent, found on an exposed attacker server. According to BleepingComputer, the framework compared software before and after patches, then wrote and ran exploits. GTIG also reported abuse of open-source tools including ARTEX and CAIRN, and expects unsophisticated attacks needing little human input within about three months.