N021: Northflank customers' AI keys were abused in October's ninth AI-platform breach

Northflank told customers to treat every stored secret as decrypted after AI keys were abused; Wiz counts nine AI developer platforms breached in October.

Share

N021 · Monday 2 November 2026 · 5-min read · Coverage: Fri 30 Oct 04:00 PT to Mon 2 Nov 04:00 PT

Jump to: Incidents · Agent threats & vulnerabilities · Policy, protocols & compliance · Models & infrastructure · Autonomy, swarms & robotics · Alignment & safety · Other big security news · Pro deep dives

Today in 30 seconds

  • Northflank, a cloud hosting platform, told every customer on 30 Oct to treat all stored environment variables as decrypted after users saw their AI provider keys abused; Trigger.dev disclosed an environment-variable breach on 1 Nov.
  • Wiz counts at least nine AI developer platforms breached in October; stolen Amazon Bedrock keys were first used within 30 minutes, and one key drove about 1.7 million model calls in 13 hours.
  • Chainalysis says attackers using "offensive agentic techniques" got into its internal tools and took customer users' names, emails and role labels.

Top story

Northflank's breach was the ninth AI-platform breach in October, and the stolen AI keys were used within 30 minutes

30 Oct–1 Nov 2026 · Northflank status page · Wiz · Trigger.dev · Incident · no CVE · Relevance 9

Attackers broke into platforms where developers host apps and agents, read the stored secrets, and spent the AI keys they found on other people's bills. Server-side secret storage failed: Northflank's environment variables, which hold the provider keys agents use, must be assumed decrypted, and users confirmed AI-key abuse. Wiz says stolen keys were tested within 30 minutes and resold.

For leaders: Long-lived AI keys on third-party hosting platforms are a known target; ask which platforms hold yours and how fast you can revoke them.

For builders: Rotate every key stored on Northflank or Trigger.dev today, then move to short-lived, scoped credentials with per-key spend alerts.

Pro deep dive ↓

Incidents

Chainalysis says attackers used "offensive agentic techniques" to get into its internal tools

30 Oct 2026 · Chainalysis · Incident · no CVE · Relevance 8

A blockchain-investigations firm says intruders using AI agents exploited flaws and copied customer users' names, emails, IDs and roles. It hasn't named the flaws and sees no access to case data, passwords or session tokens.

Do this: If you use Chainalysis, warn provisioned users about targeted phishing and unexpected MFA prompts, and enforce two-factor login.

Pro deep dive ↓

Agent threats & vulnerabilities

Update: Red Hat rates the AI-found Linux flaw critical, with no fixed kernels yet

30 Oct 2026 · Red Hat RHSB-2026-016 · Vulnerability · CVE-2026-98432 · Relevance 8 · Earlier: N019

A Linux bug that Anthropic's unreleased AI model turned into a remote root exploit is now rated critical by Red Hat, and there is no patch yet. The use-after-free in IPv4 TCP connection handling needs no login and hits RHEL 8–10 and OpenShift.

Do this: Test Red Hat's interim nftables rule on internet-facing RHEL hosts, including agent sandbox hosts; patch when kernels ship.

Pro deep dive ↓

Policy, protocols & compliance

Also:

Chart of the day

Two stolen Bedrock keys made 2.5 million model calls in under two days

Case A: one key, about 30 hours

~800,000 calls

Case B: one key, about 13 hours

~1,700,000 calls

Source: Wiz, "Black October" for AI Developer Platforms, 30 Oct 2026 (primary). Wiz estimates over $100k in usage at list price.

Models & infrastructure

Google's Gemini 4 Carbon goes first to vetted cyber defenders because of its hacking skill

28 Oct 2026 (catch-up) · Google · Model release · Closed · Relevance 7 · Capability shift

Google's newest AI model is good enough at finding and exploiting software flaws that Google is releasing it first only to approved security teams. Google reports 88.2% on SEC-bench Pro and 54.8% on ExploitGym (self-reported).

Do this: Assume attackers will get comparable models; shorten your internet-facing patch window now.

Pro deep dive ↓

Autonomy, swarms & robotics

A ransomware crew's AI framework drove 32 MCP servers and listed 62 victims

20 Oct 2026, updated 23 Oct (catch-up) · SOCRadar · Incident · no CVE · Relevance 8

Criminals let an AI agent plan and run ransomware break-ins, with a human approving each step. SOCRadar found "Phantom" on an exposed server: OpenCode driving commercial models and 32 MCP servers, with 62 victims in unverified operator records.

Do this: Patch FortiClient EMS, SimpleHelp and cPanel, and hunt for a scheduled task named RealtekAudioSvc.

Pro deep dive ↓

Alignment & safety

Also:

Other big security news

  • FBI names North Korea for the Binance theft: the FBI ties the $6 billion-plus theft to TraderTraitor and asks exchanges to block 628 addresses (FBI, 31 Oct).
  • 319 Chrome extensions hijacked searches: Socket found extensions with about 13.4 million installs routing searches through operator servers; IT should audit extensions (Malwarebytes, 30 Oct; secondary).

One concept

LLMjacking

Running model calls on stolen cloud or AI provider credentials, often to resell cheap access. Today's Bedrock cases started within 30 minutes of a key leaking.

Basic mitigation: use short-lived, narrowly scoped model credentials, restrict where they can be used from, and alert on spend and call-volume spikes per key.

Full glossary


AgentSec Brief is published by AM Technology Solutions. Researched and drafted with AI agents, verified against primary sources, and reviewed before publication. Informational only; not legal, compliance or security advice. Verify before acting.

Pro members get the full deep dives with attack diagrams, detection queries and the fix in order, plus the vulnerability tracker, compliance calendar, open-source picks and watch list. Upgrade at https://www.agentsecbrief.com/#/portal/signup