How LinuxGuard Defends Against Autonomous Offensive Frameworks (AOFs), Including ARTEX

Attacks run by AI are happening now, not in the future. Criminal groups and state-backed groups are using Autonomous Offensive Frameworks (AOFs) to break into organisations faster than human teams can respond. ARTEX is the best-known AOF but it is not the only one. In many documented cases, the attacker relied on an account that had more access than anyone realised. This article explains what AOFs are and shows, step by step, how LinuxGuard stops them on Linux servers.
What is an Autonomous Offensive Framework?
An AOF is software that does the job of a competent penetration tester automatically and at high speed. One orchestrator agent hands tasks to specialist sub-agents for reconnaissance, exploitation, privilege escalation and lateral movement. These tools were built for authorised testing. Attackers use them because they are free, capable and easy to run.
AOFs don't invent new attack methods. They run familiar techniques without a break and move to a new route as soon as one is blocked. At an emergency Cloud Security Alliance CISO Huddle on 17 November 2026, speakers and attendees reported a sharp rise in AOF use in active campaigns, across a wide range of threat actors.
The frameworks currently in use include:
- ARTEX. An open-source, multi-agent penetration testing framework released on GitHub in July 2026. Its creator withdrew it in October 2026 because so many attackers were using it, but forks keep it in circulation, including ScopeWeaver and a paid version called ARTEX PRO. Wiz counted more than 12,000 forks before it was withdrawn.
- PentAGI, HexStrike AI and CAIRN. Google Threat Intelligence Group and Mandiant have seen threat actors misuse these frameworks in real intrusions, and actors have experimented with CyberStrikeAI.
- exploit_agent. Google found this multi-agent framework on an exposed staging server. It compared software before and after security patches, found likely flaws, wrote exploits and ran them. Google described it as the first framework it had seen that automated both finding and exploiting flaws without ongoing human direction.
There are many more. GitGuardian counts about 45 other open-source agentic attack frameworks. Taking one tool offline does little. Any defence has to work against the whole category.
Why AOFs matter to any business running Linux
Speed. AOFs cut the time from first foothold to full control from hours to minutes. The FBI reports that some intrusions went from first targeting to privileged access in under three hours. At the Dutch Institute for Vulnerability Disclosure (DIVD), the attacker went from the helpdesk application's account to root in about a minute. Google recorded 27 events in 23 seconds while an agent tested what a stolen cloud key could do.
Linux service accounts are the first identity. At DIVD the way in was the helpdesk application on a Linux server. Before the attacker controlled the machine, it was acting as that application's system account, a kind of account that few teams monitor.
Blind spots are targeted. One VAPOR PANDA operator told its agents to look for Windows machines without a CrowdStrike sensor. Any server or account nobody is watching becomes a way in.
Neglect, not zero-days. Most reported intrusions didn't need a zero-day. They used forgotten keys, unused vendor integrations, open sign-up flows and secrets left in logs.
Detection alone is not prevention. Google says AOFs currently give up stealth for speed and leave traces that defenders can detect. At this speed, though, an alert that waits for a person to act arrives too late. Protection has to act on the server automatically.
What AOFs have done so far
| Actor | Framework | What happened |
|---|---|---|
| VAPOR PANDA (China-linked espionage) | ARTEX | From 17 October 2026, targeted more than 200 organisations in at least 14 countries in 11 days; CrowdStrike confirmed post-exploitation activity at nine. |
| Storm-0940 (China-aligned) | ARTEX | From September 2026. After the ConnectWise compromise, attempted to sign in to more than 2,250 Microsoft 365 tenants; an earlier password spray peaked at 2,900 failed sign-ins per minute. |
| Onyx Sleet and Pearl Sleet (North Korean) | ARTEX | Used ARTEX in operations since late October 2026. |
| Storm-3358 (Cambodia-based, criminal) | ARTEX | Targeting financial-sector organisations since October 2026. |
| TAG-217 | ARTEX | Attacks on South Korean financial institutions; tooling prepared from August 2026. |
| Criminal group in Seoul | ARTEX | In a case reported on 26 October 2026, Seoul police said seven suspects breached more than 90 organisations, stole about 8.2 million records and demanded ransoms from 38 victims. |
| DIVD, Netherlands | ARTEX | 21 September 2026: two zero-days in the Zammad helpdesk on a Linux server; root about a minute after the application account was taken, twelve minutes after first access. |
| CRAFTY FALCON (criminal) | ARTEX, PentAGI, CAIRN, CyberStrikeAI | Named in a critical financial-sector advisory from India's CERT-In on 4 November 2026. |
| Unnamed actor | PentAGI (customised) | Systematic reconnaissance and vulnerability scanning across Asia, late August to early October 2026. |
The pattern is the same whatever framework is used: find an exposed application, take over the account behind it, collect every password and key that account can reach, then move on.
How LinuxGuard stops AOFs, step by step
AOFs follow the same chain of steps whichever framework runs them. LinuxGuard stops six of ten steps and limits the other four.
| # | What the attacker does | Real example | How LinuxGuard responds | Result |
|---|---|---|---|---|
| 1 | Exploiting a known software flaw | At DIVD, the agent tried a 2019 mail-server vulnerability. | The vulnerabilities module finds vulnerable software and patches it, with approval. | Stops |
| 2 | Exploiting a zero-day flaw | The DIVD helpdesk. | File security policies block the actions the exploit depends on. | Stops, when the right policies are in place |
| 3 | Acting as the application's system account | DIVD; web shells on compromised servers. | Every action is tied to its account. Suspicious actions are blocked and the session is ended. | Limits |
| 4 | Taking full administrator control | DIVD reached root in about a minute. | Standing routes to admin, such as excess sudo rights and old SSH keys, are removed, and file policies block exploit actions. | Limits |
| 5 | Collecting stored passwords and keys | Application servers hold database passwords, email credentials and system keys. | Reads and copies of those files are blocked and the session is ended. | Stops |
| 6 | Stealing cloud credentials | Agents using leaked long-lived cloud keys. | Network rules per account block the request. | Stops |
| 7 | Creating its own access | Storm-0940's abuse of OAuth applications. | New keys and admin rights appear in LinuxGuard as they are created, and LinuxGuard Access replaces permanent access with time-limited access. | Limits |
| 8 | Moving to other systems | DIVD's attacker used the helpdesk server to reach other services. | Network rules stop connections outside policy. | Stops |
| 9 | Leaving a back door | VAPOR PANDA deployed a backdoor; exploit_agent carried scripts to install web shells. | New files written by the compromised account are detected and can be blocked. | Limits |
| 10 | Sending data out | Command results sent back to 491 rotating webhook addresses. | Allow-listed destinations block every new address. | Stops |
What AOFs teach about access on Linux
- An application's account defines the damage. At DIVD the attacker started from the helpdesk's system account, and what that account could reach decided what happened next. Most organisations can't answer that question for their own servers.
- Passwords and keys sit in files on servers. AI agents find and collect credentials in seconds, so blocking that collection is the cheapest point to defend.
- Old credentials become open doors. Barracuda disclosed an intrusion that used a valid cloud access key from a former third-party provider that had never been revoked. On Linux the same risk sits in SSH keys and admin rights nobody has reviewed.
- Attackers grant themselves access. In one case Microsoft documented, an AI agent went from finding a leaked secret on GitHub to taking data from a cloud key vault in five minutes. New keys, rights and accounts must be visible the moment they appear.
How LinuxGuard protects your Linux estate
LinuxGuard secures the accounts, keys, admin rights and stored credentials on every Linux server, along with what each one can reach. It covers containers, Kubernetes, physical servers, on-premises, cloud and hybrid environments. It is delivered as SaaS, as a dedicated cloud tenancy or as an on-premises console.
Five capabilities do the work against AOFs:
- See every account and route to admin. Users, SSH keys, admin rights and group memberships in one risk-scored map. AI agents running on your servers appear as identities, together with the credentials they can reach.
- Patch known flaws. The vulnerabilities module updates vulnerable software, with approval.
- Block on the server. File security policies stop unauthorised reads, copies and changes, including the actions a zero-day relies on. Network rules per account block connections outside policy.
- Respond in seconds. Sessions end automatically when a threat is detected. Locking an account or removing rights takes one approval, reverses automatically and is fully logged. AI-agent credentials can be revoked across all servers in one approved action, effective within 90 seconds.
- Remove standing access. LinuxGuard Access issues time-limited access. AI agents request access with tokens that last 15 minutes, behind an approval or policy gate.
Every event is tied to its account and sent to your SIEM. Controls start in monitoring mode, move to blocking when you're ready, and include a kill switch.
Three questions for your business
- If an AOF took over the account behind one of your internet-facing Linux applications today, which passwords could it read and where could it connect?
- Could you stop that account from reading credentials or connecting out within seconds, without waiting for a patch?
- Which SSH keys and admin rights on your Linux servers have gone unreviewed for a year or more?
If any answer is "we don't know", LinuxGuard's Founding Pilot is a 60-day assessment of your own estate that answers all three.
Sources
- Cloud Security Alliance – ARTEX and Autonomous Offensive Frameworks (18 Nov 2026)
- CSA Research Note – ARTEX and Its Forks: Agentic Pentest Frameworks as Attack Infrastructure
- CSA Research Note – Crafty Falcon: Weaponized Pentest Agents Hit the Financial Sector
- FBI FLASH-20261113-001 – Threat Overview of ARTEX Autonomous Penetration Testing Framework
- Google Threat Intelligence Group – ARTEX and Friends: Enter Agentic Offensive Frameworks
- CrowdStrike – VAPOR PANDA Uses AI-Driven ARTEX to Scale a Global Espionage Campaign
- Microsoft – AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise
- Microsoft – Tracking AI adoption in North Korean threat actor tradecraft
- Microsoft – ARTEX abuse expands as cybercriminals target financial sector
- Microsoft – Threat actors use agentic AI tooling to find and abuse exposed secrets
- Recorded Future – TAG-217 Targets South Korean Financial Services with ARTEX
- BleepingComputer – Former pentester used ARTEX AI to steal 8.2 million data records
- DIVD – We got hacked, here's everything we know
- Vectra AI – ARTEX Runs Known Attacks at Machine Pace
- TNGlobal – Google finds autonomous agentic AI exploitation
- LinuxGuard – The OpenAI–Hugging Face agent intrusion, replayed against host-layer identity controls (30 September 2026, updated 17 November 2026)