Under 3 hours: I read the AI pentest code the FBI flagged
Some ARTEX intrusions went from first targeting to privileged access in under three hours, according to the FBI — and the framework that did it is 93,824 lines of Go whose authorization-scope mechanism was deleted from the code. I cloned a mirror of the deleted repo and audited the guardrails: the safety boundary is 21 database rules the operator can delete.
What the FBI actually said
I read the primary source, not the coverage: FBI FLASH-20261113-001, "Threat Overview of ARTEX Autonomous Penetration Testing Framework," 13 November 2026, TLP:CLEAR, 13 pages. The numbers in it:
- Malicious actors have used ARTEX since at least August 2026, across 10 sectors: financial services, government, technology, defense, semiconductor, energy, healthcare, manufacturing, telecommunications, and legal.
- Attacks on South Korean financial institutions exfiltrated personal information from approximately 67,000 individuals.
- A supply-chain compromise of ConnectWise Cloud Backup enabled access to more than 2,250 downstream tenants' Microsoft 365 environments.
- As of early November 2026, researchers found over 1,600 publicly exposed ARTEX console instances — and the FBI notes total deployments are likely much higher behind firewalls.
- The tool gained an estimated 1,500–2,000 GitHub stars within 10 weeks of its July 2026 release. The developer ("Autumn-27") pulled the repo on 8 October 2026 after abuse reports. Forks, mirrors, and a commercial version called ARTEX PRO keep it available.
The three-hour figure is the fastest observed window, not a benchmark — "some intrusions progressing from initial targeting to privileged access in under three hours." The FBI's point is the compression: the framework "enables threat actors with minimal technical expertise to conduct complex, multi-target operations simultaneously."
What I counted in the code
I cloned a mirror of the deleted repo (acheong08/artex, last commit 9 October 2026, an English translation of the upstream) and measured it without running it — no target, no API keys, and Sophos has warned that a backdoored ARTEX Docker image is circulating, so random images are off the table. Static analysis only. What the mirror contains:
327 Go files, 93,824 lines of Go
152 .tsx + 35 .ts files (Next.js console)
153 _test.go files
go.mod: github.com/Autumn-27/norma v0.4.3 (agent framework)
The architecture matches the FBI's description exactly. It is a multi-agent platform: a planner, recon, executor, and reporter, built on the norma framework. The agent's toolset is the standard coding-agent kit — Read/Write/Edit/MultiEdit/LS/Glob/Grep/Bash — plus three bundled skills: api-recon (endpoint enumeration with harvest scripts), playwright-cli (browser automation), and scopesentry (asset import). The planner's example goal, baked into the prompt catalog, is literally "Obtain administrator access to example.com".
And the FBI's "no brain of its own" line checks out in the config. ARTEX ships no model. It requires ANTHROPIC_API_KEY or OPENAI_API_KEY, and ARTEX_LLM_BASE_URL lets the operator point the client at any OpenAI-compatible endpoint. That is the DeepSeek door: the FBI found attacker-run instances favor Chinese models, with DeepSeek Flash "dominating usage due to their balance of capabilities, relatively low cost, and lack of observable guardrails when framed for penetration testing." The code even names deepseek-v4-pro in a comment about reasoning-model token limits. The FBI's IOC list includes api.deepseek.com.
The guardrail audit: scope enforcement was removed
This is the part that justifies the article. The safety layer lives in guard/guard.go, and its own header comment says the quiet part out loud:
// Package guard implements the safety boundary layer (docs §11): an audit log,
// user-configured intercept-rule evaluation, and Observer/G5 failure attribution.
// Every tool call passes through the PreToolUse hook before executing.
// (The RoE authorization-scope mechanism was removed; a replacement may be added
// later.) Destructive/exfil gating is no longer hard-coded here — it lives in the
// DB intercept rules (seeded as ordinary [built-in] rules, so users can disable or
// delete them), evaluated via applyIntercept.
Read that again. The rules-of-engagement scope mechanism — the thing that would confine the agent to authorized targets — was removed, with a replacement that "may be added later." What remains is a set of intercept rules stored in the database, seeded as ordinary built-in rules that any operator can disable or delete. I counted the seeded deny rules: 21 (19 in the base seed, 2 more in v2).
It gets thinner. Guard.New() constructs a guard with no intercept rules at all, documented as "used for pentest tasks where the Interceptor is not yet available." So the no-guardrails configuration isn't a hack — it's a supported constructor.
To be fair, the project is adding guardrails back. The 0.3.15 changelog (7 October 2026) added a deny rule for destructive API paths — /delete, /del, /remove, /unlink, /erase, /destroy — because the old rule only caught the DELETE method while apps delete things over GET and POST. And the api-recon skill carries a scope table that prohibits SQLi, XSS, authorization bypass, brute-force, and fuzzing. But that table is prompt text, not enforcement. A prompt is a suggestion the model can ignore; a deleted intercept rule is gone.
So the tool's safety model is: the person running it behaves. That is a reasonable design for a pentest tool operated by a professional under a signed scope. It is also exactly why the FBI's warning reads the way it does — the framework doesn't need a vulnerability to be weaponized. The "vulnerability" is that the authorization boundary was a configuration option, and attackers run their own instances.
What the agents actually did
The FBI's TTP table is worth reading because it shows what "autonomous" means in practice — not one clever exploit, but fast chaining of boring ones:
- Initial recon: sub-second Azure permission checks — test blob writes/deletes, ARM deployment dry-runs — plus directory fuzzing across 500+ requests and WHOIS asset mapping.
- Initial access: unsigned SAML assertions, XXE, SSRF, anonymous file upload, hardcoded admin passwords in SPA JavaScript bundles, missing auth on web APIs, SQL injection, SSTI, IDOR, command injection.
- Privilege escalation and persistence: Kudu/SCM credential abuse, Kerberoasting, DCSync, OAuth application credential additions, web shells. One campaign enumerated over 110,000 Active Directory objects.
None of these is novel. The novelty is the loop: an agent that can run the whole chain — recon, exploit, escalate, exfiltrate — while the operator watches a dashboard. Google observed individual automated bursts lasting 2 to 25 seconds. That's the compressed detection window the FBI warns about: noisy traffic, but too fast for a human-coordinated response.
The exposed consoles
The 1,600+ exposed consoles deserve a separate note, because they are a second-order effect of the same design. The console is a full web UI with admin password auth — 8-character minimum, bcrypt hashing — so exposure is an operator choice, not a missing login page. But the codebase carries the scar of a real auth bug: a comment in server/auth.go notes that authInit once allowed "an unauthenticated request to overwrite the admin password during database errors," and handlers must now treat "unreadable" as distinct from "unset." The fix is in. The lesson is the same as the guardrail story: when the safety boundary is configuration, every operator becomes a security engineer, and 1,600 of them got it wrong enough to be countable from the internet.
What to do about it
The FBI's mitigations, condensed — these are for defenders, and they read like a list of everything ARTEX is good at finding:
- Inventory every internet-exposed application, including forgotten portals, partner systems, and supplier integrations. ARTEX's best trick is finding the peripheral system nobody owns.
- Restrict admin interfaces to trusted networks — Kudu/SCM endpoints, hypervisors, backup platforms. The ConnectWise path ran through a trusted supplier relationship.
- MFA everywhere, and rotate any credential that has ever been exposed. The campaigns leaned on stolen credentials and unsigned SAML, not zero-days.
- Watch for the traffic shape, not just the IOCs: sub-second permission probing (test blob writes/deletes), ARM deployment dry-runs, and egress to LLM API endpoints. The FBI's IOC list is historical by its own admission — IPs get reassigned.
And if you run a pentest practice: the tool you use to test your clients is now the tool in an FBI FLASH. That's not a reason to stop using autonomous tooling. It is a reason to treat the scope mechanism as a product requirement, not a config file — because the version of ARTEX the attackers run has the scope mechanism deleted, and their copy works fine.
Caveats, stated plainly
- I analyzed a mirror, not the deleted upstream. The mirror's last commit is 9 October 2026; the upstream changed after that only in ways I can't see. The numbers above are from the mirror.
- I did not execute any of it. The counts (files, lines, rules) are from the source; the campaign numbers are the FBI's, from the FLASH PDF.
- "Under three hours" is the fastest observed window, and "2,250 tenants" is the number whose environments were accessible through ConnectWise — the FBI's wording is "enabled access to," not "breached."
- The 67,000 figure is the FBI's, sourced to Korean reporting. I did not independently verify it.
Sources: FBI FLASH-20261113-001, "Threat Overview of ARTEX Autonomous Penetration Testing Framework" (13 Nov 2026); ARTEX mirror source code, cloned 17 Nov 2026 (acheong08/artex, 327 Go files / 93,824 lines, guard/guard.go, agent/provider.go, db/db.go, server/auth.go); TechCrunch, "FBI says Artex AI open source hacking tool used to breach networks in 'under three hours'" (16 Nov 2026).
Top comments (0)