Skip to main content

Agentic SOC Alliance Expands as Enterprises Prioritize Autonomous Security

Search
  • Solutionschevron right
  • Industrieschevron right
  • Platformchevron right
  • Resourceschevron right
  • Customerschevron right
  • Partnerschevron right

What is an Autonomous Attack Framework, and What Should Security Leaders Do Now?

Share blog icon

Back to top

Back to top

November 19, 2026

What is an Autonomous Attack Framework, and What Should Security Leaders Do Now?

How these frameworks work, what researchers have observed, and five priorities for your SecOps team

Security teams have long used attack frameworks to test defenses and reproduce an intruder’s behavior. Those tools already automate many tasks. Autonomous frameworks extend that capability: an AI model can interpret results, choose a next step, and adjust its approach without a person directing each action.

That capability supports authorized security testing. Attackers are also using it to carry out parts of real intrusions.

For security leaders, the practical question is how this changes their exposure and response. How could an attacker use an initial foothold to reach sensitive systems or data? What would limit the attacker’s access? And could the team detect and contain the activity before it reaches a critical system?

DEFINITION
Autonomous attack framework – Software that uses a large language model (LLM) and connected tools to plan and carry out attack steps, evaluate results, and decide what to try next without a person directing every step. It can support authorized security testing such as adversary simulation and detection testing, or be misused by attackers.

What makes an attack framework autonomous?

An attack framework provides reusable tools for carrying out and coordinating offensive security techniques. During authorized testing, it helps teams identify exploitable weaknesses, investigate what access they provide, and collect evidence. Metasploit, for example, supports these activities as part of a structured testing workflow.

An autonomous framework connects an LLM to those kinds of tools and lets it work toward an objective. The model helps decide what to do, then the tools perform the actions, and the framework feeds results back into the next decision.

How an autonomous attack framework works

An agent might discover a service (e.g., a web application, a database, a remote access service such as SSH), investigate its configuration, attempt an action, and use the response to decide what to try next. The additional capability is delegating more of the interpretation and decision-making, including troubleshooting failures.

“Autonomous” does not mean there is no human involvement. A person may still choose targets, supply access, approve important actions, or intervene when the system gets stuck. Anthropic’s account of an espionage campaign illustrates this division of work between an attack framework and human operators.

What has actually been observed?

These examples show how attackers are using AI to carry out intrusion tasks, act on exposed credentials, and turn software patches into working exploits.

  • Delegating substantial portions of an intrusion. In November 2025, Anthropic reported a campaign targeting approximately 30 organizations. It estimated that AI performed 80–90% of the operational work, with humans retaining key decisions. A small number of intrusions succeeded. The company also observed mistakes, including invented credentials and inaccurate claims of success. Those figures describe Anthropic’s assessment of one campaign.
  • Moving quickly from exposed credentials to data access. In one observed case, Microsoft reported an agent progressing from searching GitHub for exposed secrets to taking data from Azure Key Vault in five minutes. In related activity, failed cloud requests were followed by corrected requests within seconds.
  • Using patches to develop exploits. In November 2026, Google reported a framework called exploit_agent that compares software versions around patch announcements, identifies potential vulnerabilities, develops exploits, and executes them automatically. This illustrates how automation can shorten the interval between disclosure and exploitation.

These reports establish that attackers are using the capability. They do not establish that every attack is autonomous, that every agent succeeds, or that all organizations face identical exposure.

What does this change for security leaders?

It makes it more important to understand how one compromised application or credential could provide access to other systems.

A compromised application may have permission to read a credential store. A credential found there may provide access to another system. An agent can investigate those connections and attempt several paths in rapid succession.

Trusted third-party access deserves attention here. Microsoft reported that attackers used ARTEX to attempt sign-ins across more than 2,250 tenants after compromising ConnectWise application identities. It also cautioned that successful authentication did not necessarily indicate data theft. The practical question is how much access a supplier’s application has, and how your organization would withdraw it.

There are opportunities for defenders, too. Google reports that current frameworks can produce conspicuous activity and recognizable traces, sometimes sacrificing stealth for speed. Their limitations matter when assessing the threat.

Five practical priorities

1. Identify exposed systems

Ask for a current view of internet-facing applications, administrative interfaces, and their owners. Prioritize remediation using exposure, evidence of exploitation, business importance, and the access a compromise would provide.

For urgent vulnerabilities, establish a clear path to patching, restricting access, or temporarily isolating an affected service. Include applications and dependencies in that process.

Ask your team: “How quickly can we identify affected systems and put a mitigation in place?”

2. Limit what software identities can reach

Applications, service accounts, and automated workflows have identities and permissions of their own. Prioritize those that can administer systems, retrieve credentials, or reach sensitive data.

Give them accountable owners, remove unnecessary permissions, and replace persistent secrets with temporary or workload-bound credentials where supported. Restrict where they can authenticate as well as what they can access.

Continue scanning for exposed credentials. Deleting a secret from a repository does not invalidate it; revoke or rotate it and investigate its use.

Ask your team: “Which software identities would cause the most damage if compromised?”

3. Detect suspicious behavior

Look for what an attacker is doing, including when they use legitimate credentials and familiar tools. Useful signals include an application identity accessing unusual resources, unexpected permission changes, or a rapid sequence of credential access followed by large data downloads.

Context matters. A burst of activity may be normal for a backup service and unusual for an application that typically reads a few records. Establish expected behavior, then investigate deviations across identity, cloud, endpoint, and network activity. Mandiant emphasizes this combination of identity, origin, volume, and sequence in its detection guidance.

These detections do not need to establish whether AI is involved. They need to identify suspicious activity early enough for the team to act. Test that alerts reach the right responders with enough context to make a containment decision.

Ask your team: “If someone used a valid application credential to explore our environment and access sensitive data, what would alert us?”

4. Practice containment

Agree in advance on the evidence and authority needed to disable an identity, restrict access, or isolate a workload. Define which actions can be automated, which need human approval, and how to recover from a mistaken intervention.

A sensible first step is to automate well-understood actions with bounded consequences, then test them. Faster response does not require giving a system unrestricted authority.

Also verify what containment actually accomplishes. Already-issued tokens can remain valid until expiry. Changing a credential and ending existing access are not always the same operation.

Ask your team: “What can we safely stop now, and how do we verify that access has ended?”

5. Secure your own agent tools

Your organization’s coding assistants and business agents introduce a related security responsibility: controlling the software, configuration, and credentials they use.

In November 2026, malicious releases of Microsoft’s Python agent-framework package carried a credential stealer. This was a software supply-chain compromise affecting agent-building software.

Include agent tools in dependency review, change control, access management, and monitoring. Review executable extensions and configuration changes, and avoid giving development environments broad production access by default.

Ask your team: “Who owns these tools, what can they execute, and which credentials can they reach?”

A useful place to start

At your next security leadership review, choose one critical service. Ask the application owner, identity team, and security operations team to walk through a compromise together:

  • What would the attacker initially control?
  • What could that access lead to?
  • Which activity would trigger investigation?
  • Who could contain it, and how would they confirm containment?

Leave that discussion with a short list of gaps, an owner for each, and a plan to test whether the fixes work.

Frequently asked questions

What is ARTEX?

ARTEX is an open-source, multi-agent penetration-testing framework that can use any AI model to run an attack from reconnaissance through data exfiltration with little human involvement. Microsoft has reported its use by Storm-0940, Storm-3358, and other actors in 2026. Learn how it works.

Is an autonomous attack framework the same as AI-written malware?

No. These frameworks mostly drive ordinary, well-known tools such as Nmap, curl and cloud APIs. The AI decides what to do next and does it faster and more persistently than a human.

Can we block ARTEX by its IP addresses or user agents?

Only briefly. Microsoft notes that agents change these fingerprints quickly, so behavioral detection on identities and cloud APIs matters more than static blocklists.

Was the Microsoft agent-framework compromise caused by AI?

No. It was a software supply chain attack on a package used to build AI agents. It matters because those packages run where valuable cloud and CI/CD credentials live.

How fast can an AI-driven attack move?

Microsoft has reported intrusions that reached privileged cloud access in under 3 hours, and one agent that went from finding a leaked secret on GitHub to stealing Key Vault data in 5 minutes.

What is the single most important control?

Strong governance of non-human identities: knowing every app and service credential you have, giving each the least privilege it needs, replacing long-lived secrets, and being able to disable any of them in minutes.

blog image
Robyn Fisher
Robyn Fisher

Principal Product Marketing Manager

Robyn Fisher is a Principal Product Marketing Manager at ExtraHop focused on AI, cybersecurity, and the evolution of security operations. She combines customer research, engineering collaboration, and analysis of emerging threats to help security leaders understand how new technologies are changing cyber defense. Previously, she held product marketing roles at Google, Amazon, Microsoft, and Volkswagen Automotive Cloud. View her profile on LinkedIn.

Share
LinkedIn logoX logoFacebook logo
Key Takeaways
  • Autonomous attack frameworks use LLMs and connected tools to carry out attack steps and adapt without a person directing every action
  • They support authorized security testing, but attackers also use them to accelerate parts of real intrusions
  • Practical priorities include: reduce exposure, limit software identities’ access, detect suspicious behavior, practice containment, and secure internal agent tools

Explore related articles

Experience RevealX NDR for Yourself

Schedule a demo