N024: OpenAI's agents wrote 948 working Linux root exploits in 30 hours

OpenAI's agents produced 948 working Linux root exploits, Claude found a Xen VM-to-VM bug, and GitBook blames AI agents for a breach: the walls around agent code are getting cheaper to break.

Share

N024 · Thursday 5 November 2026 · 5-min read · Covers 4 Nov 4 AM to 5 Nov 4 AM Pacific

Jump to: Incidents · Agent threats & vulnerabilities · Policy, protocols & compliance · Architecture & harnesses · Autonomy, swarms & robotics · Other big security news · Pro deep dives

Today in 30 seconds

  • OpenAI's agent pipeline turned 2,308 Linux kernel findings into 948 working root exploits in 30 hours, so a container whose only wall is the shared host kernel no longer reliably holds agent-run code.
  • Anthropic's Claude found a Xen bug (XSA-523) that lets one guest VM read and write other VMs' memory, the next wall many agent sandboxes rely on.
  • GitBook says an attacker it believes used AI agents turned a flaw in GitBook's SSO login into staff sessions and production secrets in under three hours.

Top story

OpenAI's agents turned 2,308 Linux kernel findings into 948 working root exploits in 30 hours

4 Nov 2026 · OpenAI · Research · no CVEs listed · Capability shift · Relevance 9

An OpenAI system found hundreds of ways for an ordinary program on a Linux machine to take full control of it, and that machine is usually the wall around an AI agent's code. OpenAI says its five-stage agent pipeline, running an unreleased cyber model with human review before reporting, reproduced 1,724 kernel bugs and produced 948 working local root exploits for about $1,280 each in API credits. It hasn't said how many are fixed.

New risk for leaders: shared-kernel container sandboxes now face root bugs found at machine scale, faster than patching alone can absorb.

For leaders: Ask which agent sandboxes rely only on the shared Linux kernel, and fund moving untrusted agent code behind a second kernel boundary this quarter.

For builders: Run agent-generated code under gVisor, Kata or Firecracker instead of plain containers, and keep credentials off sandbox hosts.

Pro deep dive ↓

Incidents

GitBook says an attacker using AI agents reached production 76 minutes after gaining staff sessions

3 Nov 2026 · GitBook incident report · Incident · no CVE · Relevance 9

A documentation platform was broken into, and the attacker reached customers' connected secrets within hours. GitBook says its SSO login flaw let the attacker act as staff, then a misconfigured GitHub App and a Cloudflare key led to production; the AI-agent claim is GitBook's inference from the speed.

Do this: GitBook users: issue new API tokens, rotate integration secrets, re-authenticate its MCP server.

Pro deep dive ↓

Agent threats & vulnerabilities

Infostealers now copy Claude Code and Codex logins; a password reset doesn't revoke them

30 Oct 2026 (catch-up) · SpyCloud · Threat research · no CVE · Relevance 8

Password-stealing malware on developers' PCs now copies the files that keep AI coding assistants signed in, so criminals can use or resell them. SpyCloud says two stealers take Claude Code's OAuth tokens, including MCP server tokens, and Codex CLI's API key.

Do this: After an infection, revoke Claude Code and Codex sign-ins and rotate AI keys from a clean machine.

Pro deep dive ↓

Also:

Policy, protocols & compliance

Also:

Chart of the day

One 30-hour run: 948 working root exploits

Candidate kernel findings

2,308

Bugs reproduced

1,724

Working root (LPE) exploits

948

Info-leak proofs of concept

291

Source: OpenAI (4 Nov 2026; run in mid-October), via The Hacker News (secondary).

Architecture & harnesses

Claude found a Xen bug that lets one VM read and write another VM's memory (XSA-523)

3 Nov 2026 · Xen advisory, via The Hacker News · Vulnerability · CVE-2026-102703, CVE-2026-102704 · Relevance 8

A flaw in Xen, software that splits one server into many virtual machines, lets a malicious machine read and change its neighbours' memory. Xen's grant-copy code skipped a permission recheck; Xen 4.6 and later are affected and no attacks are known.

Do this: Patch and reboot Xen hosts running agent sandboxes; ask sandbox vendors if they use Xen.

Pro deep dive ↓

Autonomy, swarms & robotics

A one-person, 12-agent attack mesh switched to safety-stripped models when commercial ones refused

4 Nov 2026 · Acronis TRU · Threat research · Asia-Pacific · Relevance 8

One person in Taiwan used a team of AI agents against emergency services and universities and stole medical data. Acronis says commercial models refused, so the main attack agent ran on safety-stripped local models; about 10 compromises are confirmed.

What it means for US companies: refusals didn't stop it; closing login-free services is the control.

Do this: Close internet-facing services that need no login or use default credentials.

Pro deep dive ↓

Other big security news

  • Apple flaw on CISA KEV: APT29-linked attacks use CVE-2024-27834, fixed May 2024; update old iOS, macOS, Safari (CISA) (actively exploited).
  • Rails upload flaw, CVSS 9.8: apps using ImageMagick for uploads let anyone write server files; upgrade to 8.1.4.1 or 8.0.5.2 (GHSA-f3w6-w445-fwh9).
  • MUJI card data: about 1.09 million card numbers may be exposed; watch statements (Tokyo Reporter, secondary).

One concept

Local privilege escalation (LPE)

A bug that lets a program running with limited rights take full administrator ("root") control. Agent code usually runs with limited rights in a container sharing the host's kernel, so one kernel LPE can hand over the host.

Basic mitigation: Add a second kernel boundary (gVisor, Kata or Firecracker) and patch hosts on a short cycle.

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