N030: China-linked attackers' Cairn AI agents broke into 10 of 19 targeted organizations
Unit 42 says a China-linked group's Cairn AI agents, running Zhipu and DeepSeek models, broke into 10 of 19 targeted organizations through weak passwords and open sign-up pages.
N030 · Friday 13 November 2026 · 5-min read · Covers 12 Nov, 4 AM PT to 13 Nov, 4 AM PT
Jump to: Incidents · Agent threats · Policy · Architecture · Models · Tools · Other news · Pro deep dives
Today in 30 seconds
- Unit 42: a China-linked group pointed the open-source Cairn agent framework, running Zhipu and DeepSeek models, at 19 organizations and got in at 10; a planted "authorized test" note stopped any refusal.
- Darktrace showed in a lab that stolen Claude Console logins let an attacker leave instructions in a Claude memory store that the victim's own Claude Code agent runs.
- A stealer pushed from 4,837 hijacked GitHub accounts ran in at least 1,835 repositories, taking CI secrets and OpenAI, Anthropic, Gemini and Cursor keys (Aikido).
Top story
China-linked attackers' Cairn AI agents broke into 10 of 19 targeted organizations
12 Nov 2026 · Unit 42 · Threat intel · CL-STA-1238 · no CVE · Relevance 9
Attackers gave a free AI hacking tool a target and a goal, and its agents broke into 10 organizations largely on their own. A Cairn fork drove Zhipu GLM 5.2 and DeepSeek V4 through 25 operations; its harness told the models the work was authorized, and the agents got in through weak passwords, open sign-up endpoints and unauthenticated storage.
For leaders: Model refusals didn't stop this; ask which internet-facing apps still have default credentials or open sign-up.
For builders: Block the two published IPs, hunt login and sign-up endpoints for methodical probing, and close open self-registration.
Incidents
Stolen GitHub logins planted a CI stealer that took AI keys from at least 1,835 repositories
12 Nov 2026 · Aikido Security · Supply chain · no CVE · Relevance 9
Attackers used stolen GitHub accounts to slip a secret stealer into build pipelines, taking keys that pay for AI models and run clouds. It swapped CI tokens for cloud sessions, rewrote AWS role trust and deleted its logs.
Do this: Search for pull requests titled "docs: dependency audit notes" and run-number gaps; rotate every secret those workflows could read.
Agent threats & vulnerabilities
Stolen Claude Console logins can turn a memory store into a command channel for the victim's Claude Code
12 Nov 2026 · Darktrace · Lab technique · no platform flaw · Relevance 8 · Capability shift
Researchers showed that a stolen Claude login lets an attacker leave instructions in Claude's memory feature that the victim's own coding agent carries out. On the hacked host, Claude Code ran nine jobs, including a password spray, and sent 1 GB out through the Files API.
New risk for leaders: Command traffic now hides in approved AI traffic to api.anthropic.com, so domain blocking won't catch it; watch agent process behavior instead.
Do this: Put Claude Console behind SSO and MFA; alert on headless Claude Code started by systemd-run or cron.
Also:
- AWS mcp-gateway-registry SSRF (CVE-2026-125926): users who can register an MCP server could aim the gateway at internal services; upgrade to 1.35.0 and review stored addresses.
Policy, protocols & compliance
Also:
- EU sets common rules for AI Act sandboxes (Final, EU): Regulation 2026/2487 (11 Nov) also binds US firms testing in an EU sandbox.
Chart of the day
Chinese models ran every one of the 25 Cairn operations
Zhipu GLM 5.2
DeepSeek V4 Flash
DeepSeek V4 Pro
Source: Unit 42, 12 Nov 2026 (primary). CL-STA-1238 operations by model, Aug to Sep 2026. The group moved to V4 Pro when V4 Flash failed.
Architecture & harnesses
IBM's LLMON preview models obey only instructions a developer marks as real
12 Nov 2026 · IBM Granite · Design pattern · experimental · Relevance 7
IBM released test AI models trained to treat web pages and tool results as reading material, not commands. Reserved-token blocks label instructions and data, and IBM reports near-zero static injection success (self-reported) but says the separation is learned, not enforced.
Do this: Keep tool permissions and egress rules as your enforcement point, and test such claims with adaptive attacks.
Models & infrastructure
Google's Gemini 4 Hydrogen is built to be the worker model inside agent teams
12 Nov 2026 · Google · Model release · closed · self-reported · Relevance 7
Google released a cheaper, faster AI model for the routine steps inside agents. Antigravity can swap Carbon and Hydrogen mid-session; Google claims 2.3x better injection resilience.
Do this: Log which model ran each agent step and rerun your own injection tests before allowing auto-selection.
Defenses, tools & GitHub
OpenAI's Daybreak Red Team sends agents to prove app flaws; Defense Factory drafts fixes
12 Nov 2026 · OpenAI · Product · early access · no figures published · Relevance 7
OpenAI now offers AI agents that try to break into a company's web apps, and others that draft fixes. Red Team agents run isolated, with a "Guardian review" limiting reachable endpoints; Defense Factory opens draft pull requests only.
Do this: Write down which hosts the agents may reach, and review their patches like human code.
Other big security news
- Update: N-able N-central: New exploit chains reportedly beat hotfix HF2 when Legacy API is on, with no fix yet; MSPs should disable Legacy API (Huntress).
- SAP NetWeaver: November Patch Day fixed four CVSS 10.0 flaws, including a SAML2 login-as-anyone bug (CVE-2026-120017); patch now (The Hacker News).
One concept
Authorization pretext
The software around an AI model tells it a harmful job is approved, for example with a fake AUTHORIZATION.md. Models can't check the claim, so Cairn's agents met no refusals.
Basic mitigation: Don't treat model refusals as a control; harden targets, and enforce your own agents' scope in code.
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