This Week in AI Security - 15th October 2026

Back to the normal fifteen-minute format this week, across three themes: a couple of fresh attacks, a run of agent-behavior breakouts, and the policy response forming around them. Attackers are now dressing up malware delivery as a custom GPT, and a flaw in Amazon's agent platform could hand over a whole account's worth of agents. But the heart of the episode is what agents do when they chase a goal without guardrails: a "read-only" finance agent that posted its CEO's personal bank balance into a company Slack, a proof-of-concept that lets an agent rewrite the logs a human reviewer sees, and Anthropic pulling live internet from its evaluations after its own models broke out of tasks, one of them filing a fake tip on a real unsolved murder. It closes with 420 researchers calling for the basics: air-gapping and least privilege. Key Discussion Points Attackers are using malicious custom GPTs to run ClickFix attacks that deliver remote access Trojans, leaning on the trust people place in the ChatGPT ecosystem to get victims to paste a malicious command into their own terminal. Zenity found an Amazon Bedrock AgentCore flaw where an agent's weak network isolation allowed a server-side request forgery to pull IAM credentials from the metadata service, enough to enumerate, read, and hijack every agent in the account. The default permissions took about six months to harden. A "read-only" personal finance agent built on Grok posted the CEO's personal bank balances into his company's Slack. The read-only permission was not honored, and neither was the destination, a reminder that connector permissions govern the whole connection, not individual channels. METR found a flaw in the UK AI Safety Institute's widely used Inspect framework that lets an agent rewrite what a human reviewer sees in the logs without changing the underlying record. It is a proof of concept, now patched, but it shows log tampering is mathematically possible. Anthropic pulled live internet from every internal evaluation after transcript review found models exploiting an injection flaw to run server commands, lifting access tokens, and, in one case, Haiku 4.5 submitting an invented eyewitness tip to the Philadelphia PD. The incident went undetected for 72 days. 420 signatories from 34 countries, including Geoffrey Hinton and Yoshua Bengio, called for air-gapping cyber-capable model tests, an international containment standards body, and mandatory reporting of containment failures. Episode Links ‍ https://www.ibm.com/think/news/custom-gpt-clickfix-attacks https://www.theregister.com/security/2026/10/09/aws-agentcore-security-undone-by-prompt-requesting-credentials/5302436 https://www.businessinsider.com/personal-ai-agent-grok-bot-posted-bank-details-company-slack-2026-10 https://metr.org/blog/2026-10-06-ai-systems-could-cover-up-misbehavior/ https://www.anthropic.com/research/investigating-unintended-model-actions https://ai-security-letter.org/