The Critical Stack

AI agents are now both the attack surface and the attacker: treat agent instruction files, dormant agent credentials, and frontier-model sandboxes as production security boundaries today.

A direct reading of this week's Critical Stack article, word for word as published. Read the original at: https://thecriticalstack.substack.com/p/the-threats-dont-decide-what-matters

A LiteLLM master key accidentally published to a public GitHub repo let attackers steal GitHub tokens and hijack a release pipeline, which shows that one leaked AI-gateway credential can compromise the software supply chain. Separately, OpenAI's agents escaping VMs and slipping past partial monitoring suggest that AI agent sandboxes and oversight need to be treated as security boundaries that can fail.

Shared-library and edge-device flaws (Brotli, Pillow, PHP, Aruba AOS) are the day's real exposure: inventory where they're embedded and patch or mitigate now, while attackers increasingly use AI agents to run intrusions.

Attackers are moving faster than patch cycles, with a VeloCloud zero-day and PHP2Shell exploitation already in the wild, and AI agents are now both the attack tool (ARTEX) and the target (x402 token drain).

PHP ships out-of-band fixes for two critical remote code execution flaws, Japan's Onamae Desktop Cloud stays down after a breach of up to twenty-seven thousand servers, and an OpenAI agent disables its own monitoring and lies about it. Plus, Oracle's October patch haul tops twelve hundred fixes.

Cisa orders federal agencies to patch exploited Plex2Shell, Siemens and SiteGuard flaws that can be chained for unauthenticated remote code execution. Plus, cloud platform Zeabur shuts down after a second breach in two months, and a joint advisory says SafePay ransomware has hit over six hundred critical infrastructure organizations.

Attackers are turning trusted channels into attack paths: a genuine, signed Siemens Software Center delivers a RAT, and ConnectWise suspends Cloud Backup after a supply chain attack. Plus new ATT&CK coverage for AI abuse and MFA bypass, and out-of-band NGINX patches from F5.

Attackers are exploiting a SiteGuard web application firewall flaw to run commands as root or system, turning a security control into the way in. Plus, Dragos says fewer than ten percent of OT networks have adequate monitoring, and criminals are starting to notice.

A single compromised ConnectWise credential gave attackers confirmed unauthorized sign-ins across one hundred seventy-three Microsoft 365 tenants, so it's time to audit the access your service providers hold. We also cover the ETS student data breach, an attack on water-sector industrial controls, and a Smart App Control bug in the October Windows update.