Agentic AI attacks use AI agents inside attack frameworks to choose tools, read results and retry on their own. In 2026, MOXFIVE handled several intrusions suspected to use agentic frameworks, most notably ARTEX, a penetration testing tool published on GitHub in July 2026. Agents lower the skill needed, speed each attack and let one operator run many.
MOXFIVE has handled several cases that we have high confidence were carried out with agentic attack frameworks, in which AI agents choose, run and chain dozens or hundreds of tools on their own to carry out a cyber intrusion. The most notable of these is ARTEX, which was published on GitHub in July 2026 as a legitimate penetration testing tool.
Attackers have abused legitimate penetration testing tools for decades. ARTEX and similar frameworks are different: they are built so that AI agents can operate dozens, or hundreds, of tools within a single framework. Putting AI agents into attack frameworks has dramatically lowered the skill and experience needed to carry out a cyber intrusion. It has also increased how many attacks one attacker can run and how fast each attack moves.
What do agentic attack frameworks change?
- Agentic frameworks lower the barrier to entry
An intrusion used to require an operator who knew which tool to run next, how to read its output and what to do when it failed. An agentic framework handles those decisions. The operator sets a goal, and the agent picks tools, interprets results and retries until something works.
The techniques we have seen are not advanced. In one recent case, the attacker created accounts through free-trial sign-ups or used password spraying, then exploited simple flaws available after login, such as server-side request forgery. The tradecraft is basic. The difference is that the agent never runs out of patience, never runs out of ideas to try, and needs no expert behind it.
The result also looks different to investigators. In AI agent attacks, activity wanders. It probes, backtracks and tries things a human operator would skip, and it does so loudly. That noise is useful for detection, but it makes reconstructing an intrusion slower and more confusing.
- Agentic AI attacks move faster
Agents do not stop to think between steps. When something blocks them, the next attempt follows within seconds, usually adjusted to the specific error that stopped them.
In an intrusion against an AI company that we assessed as AI-assisted, the attacker created a new branch 2.6 seconds after branch protection blocked a change. It sent a rewritten payload 30 seconds after a failed code-execution attempt. It went from first activity to publishing the company's secrets store in a public repository in 15 minutes. Public reporting describes the same pattern: Microsoft documented an agentic intrusion that went from secrets exposed on GitHub to data stolen from Azure Key Vault in about five minutes.
A response process that starts when a person notices the first alert may already be too slow. Cloud audit logs often take around five minutes to arrive, and five to fifteen minutes from action to alert is common. An attack that finishes in a minute or two can be over before the first alert fires.
- One attacker can run many attacks in succession
Because the agent does the work, one operator can run many intrusions at once or back to back. Reporting by BleepingComputer and others links ARTEX to automated attacks against more than 200 organizations, by both Chinese state actors and criminal groups.
Our cases reflect this. In one environment, a single source sent 475 requests to sign-in pages in under two hours, all with the user agent ARTEX-recon/1.0. That was reconnaissance, set apart from the main incident under investigation. In another case, investigators matched the operator's TLS client fingerprint across changing IP addresses and tied the activity together even as its infrastructure rotated. Tooling that runs at this volume against this many targets means more organizations will see agentic activity, whether or not they are specific targets.
How can defenders stop agentic AI attacks?
Agentic frameworks don't need new defenses. They punish the gaps in existing ones faster and more often. Three actions reduce the most risk now.
- Enable phishing-resistant MFA
Password spraying and credential reuse are still among the most common ways in, and agents run both without getting tired. Require MFA everywhere, and use phishing-resistant methods such as FIDO2 security keys or passkeys for administrators, developers and anyone with access to cloud consoles, code repositories or secrets. Require callback verification before any MFA reset, because the help desk is the next target once MFA is in place.
- Reduce internet exposure, especially for privileged accounts
An agent can only attack what it can reach. Inventory every internet-facing sign-in page, admin portal, API and trial or self-registration flow, and remove anything that doesn't need to be public. Restrict privileged accounts and administrative interfaces to managed devices, known networks or a VPN through conditional access.
This control proved its worth in the AI company intrusion. Weeks after its last successful activity, the attacker was still trying to log in every day with valid stolen credentials. Network restrictions blocked those attempts. Credential revocation alone did not.
- Monitor the control plane for agentic activity
Agentic attackers rarely bring malware. They work through identity providers, cloud consoles, SaaS admin settings, CI/CD pipelines and secrets stores, the systems that control everything else. Endpoint tools see little of that activity.
Centralize and keep at least 90 days of identity provider, cloud audit, SaaS admin and source-control logs. Alert on the behaviors agents produce:
- Bursts of failed actions followed by quick success.
- Sudden spikes in denied API calls or secret reads.
- New tokens, keys or service-account credentials.
- Changes to build workflows.
- Activity from new trial or self-registered accounts.
These patterns are noisy and can be detected. In one of our cases, the attacker was denied direct access to secrets more than 2,600 times before finding a path that worked. Each of those denials was a chance to catch it.
The bottom line
ARTEX and frameworks like it don't introduce new attack techniques. They give old techniques to more people and run them faster, against more targets. Defenders who close the basic gaps, watch the control plane and can act within seconds rather than days will be ready for this wave.