- November 16, 2026
LARA: What an AI Attack Harness Wrote Down
Beazley Security Labs recovered memory files from a multi-agent AI attack harness that called itself LARA. The files show three agents, about 270 attack scripts, and a task-code system for tracking objectives, findings and stolen credentials. LARA tried roughly a dozen CVE exploits, and all of them failed against patched systems. It got in through misconfigurations instead: an unencrypted ASP.NET ViewState that leaked a service account credential, and a weak SFTP password. It then dropped a PHP webshell and took source code, private keys and databases. MFA, limited password reuse and network segmentation kept it from spreading further.
Executive Summary
In September 2026, Beazley Security investigated an incident and quickly determined the attack was being conducted by a multi-agent AI harness. During the investigation, the team recovered a portion of the harness’s memory (.md) files.
Study of these files reveal a capable, highly organized AI system that could carry out nearly every step of an attack chain, from reconnaissance and exploitation to credential theft, persistence, pivoting, and data exfiltration. Unlike most public accounts of AI-enabled attacks, this case offers a first-hand record of an intrusion as it unfolded, showing how the agents tracked their progress, coordinated with one another, and adapted to the defenses they encountered.
Most importantly, the logs show which of the victim’s defenses blunted the harness’ efforts and reduced the blast radius, offering practical lessons for any organization hardening against agent-driven attacks.
AI agents can perform most actions of an intrusion. The harness carried out nearly every stage of the attack chain with little human direction, coordinating across three servers and multiple agents.
The agents’ own records tell the story. We recovered memory files that show agents tracking objectives, tasks, findings, and stolen credentials through a structured coding system.
Foundational defenses held where applied. Strong patch management caused every exploit attempt against known CVEs to fail, while deployed identity controls and network segmentation limited lateral movement.
Vulnerabilities in custom applications can be found and abused agentically. The harness gained access through a custom web application and misconfigurations.
Machine speed demands faster response. Agent-driven attacks shrink the time defenders have to contain an incident and reduce impact.
Offensive AI Harness Overview
Much of the conversation around offensive AI in the hands of real-world threat actors has been speculative in recent months. The harness recovered in this investigation moves that discussion from theory to reality, and its architecture illustrates an agent-driven attack that was built, deployed, and operated against a live target.
The multi-agent harness was setup across three servers, consisting of a central “operator” host that contained all the harness objects needed to coordinate the agents, and two support servers (agents) to assist with scanning, exploit delivery, and payload staging.
Those two support servers, internally named “attack-hand1” and “attack-hand2”, appear to be configured with the same scanning, exploit, and C2 hosting capabilities as the central operation box. They were activated mostly to check and verify if the target had either blocked the central operations box or if connectivity for a particular targeted server or asset was lost.
Figure 1: Harness setup - servers
The harness’ primary mechanism for coordinating activity between agents was through a task management system with assigned “task codes” to track goals, tasks, processes, and assets during the attack. The attack was segmented into attacker defined tasks which appear to form part of the attacker’s playbook and fall into two groups: four prioritized campaign objectives, and a tracking schema for operational progress. The “task codes” were documented as follows:
Code | Category | Description |
|---|---|---|
P1 | Priority | Remote code execution (RCE) on a target host |
P2 | Priority | Financial system access |
P3 | Priority | High severity data leak |
P4 | Priority | Internal network access |
T-xxx | Task | Discrete action taken toward a priority goal |
F-xxx | Finding | Result of an action that warranted follow-up |
C-xxx | Credential | Stolen authentication material that could facilitate pivoting and persistence |
D-xxx | Done/Dead-end | Task group that was either completed or dead-ended |
At the onset of the attack, three hosts began systematically performing reconnaissance against the target network and attempted to land exploits on relevant, discovered assets. Tasks and findings were assigned IDs according to the schema above, and results were captured in one of three memory files to prevent duplicated work. The harness interestingly referred to its own memory setup as “tripod memory” consisting of the following three files:
networkmap.md : notes on discovered network hosts in the target environment
bugtrack.md : a list of vulnerable findings for software on the target hosts
loot.md : stolen credential material from any successful exploits against the target hosts
Figure 2: Offensive AI Harness “tripod memory”
Forensic analysis indicates the harness had access to an extensive library of exploit and attack packages. In addition, many of the scripts appeared to be purpose-built for this operation, or came from easily modified boilerplate, as targets were hardcoded and specific to this victim. The harness also leveraged at least one open-source repository: a FortiGate exploit kit found on GitHub.
Below is a representative sampling of the ~270 attack scripts discovered during our investigation, demonstrating capabilities targeting Azure, CrushFTP, EMS Database Management, and Fortinet software.

Figure 3: A Sampling of the Harness’ Exploit and Attack Tools
As for the name, the recovered files contain a few entries in which the harness appears to reference itself as “LARA.” Although styled in all-caps like an acronym, we could not find any entries expanding the acronym or explaining the origin.

Figure 4: “LARA”
Attack Chain(s)
With this three-agent setup, shared exploit kits, and memory files for coordination, the harness began with a reconnaissance phase, discovering and inventorying as much of the organization’s externally exposed infrastructure as possible. Using this inventory, it then deployed exploits from its library against known and suspected CVEs affecting the identified software.
Figure 5: Distribution of Exploit Attempts Across Harness Agents
The sub-agent named “attack-hand1” launched most of the exploits, however all of attacks were distributed across all three of the agents, each launching at least one attack during the campaign.
According to the recovered memory files, approximately a dozen exploits were launched, with all of them logged as unsuccessful from the agents. The resiliency appears to be thanks to disciplined patch management, reinforced by strong mitigations such as disabling unnecessary services and employing strict network segmentation. Below are screenshots from log files leveraged during the forensic investigation showcasing exploit attempts:

Figure 6: AI Logged Exploit Attempts
The agents attempted to exploit a broad range of known flaws in products visible on the target infrastructure, including Veeam, Ivanti, GlobalProtect, and CrushFTP. The targeted CVEs spanned an extensive timeframe, ranging from vulnerabilities reported back in 2019 to some discovered this year. Although the harness appears to have been unable to breach this target via CVE exploits, it pivoted and found success hunting for system misconfigurations.
One of the crucial “misconfiguration” findings discovered in this attack belonged to a vulnerable custom Active Server Pages (ASP.net) webapp. The webapp made use of a commonly targeted ViewState mechanism, which allows a web server to pass encrypted state information to a browser that persists session data between pages. The harness discovered that for one target webapp, this ViewState functionality was not configured to use encryption, and contained credentials for an internal system service account.

Figure 7: The ViewState Finding
Additional harness log entries recovered from the ViewState investigation illustrate how the “task codes” documented above were used to log the state of each task and coordinate activity between agents.

Figure 8: Harness Activity for ViewState Finding “F-087”
The harness logs illustrated above show that:
- 1.
The ViewState finding was assigned task code F-087 (F – Findings).
- 2.
Exploiting it worked towards two coded objectives: P3 (High severity data leakage) and P4 (Internal network access).
- 3.
As part of the F-087 finding,
- 4.
tasks T-007 and T-011 (portal + credential conversion attempts) were created.
- 5.
These tasks then tested the stolen ViewState service credential on other target assets.
- 6.
Once attempted, those tasks were marked “done” with task code D-033 (Done/Dead-End).
How the harness handled the ViewState finding reveals other interesting behaviors worth investigating. First, the harness took steps to setup a monitoring system to watch the service account and notify when it saw the password change.

Figure 9: Monitoring Password Changes for the Visible Service Account
Once the harness was able to pivot into other systems, any credentials found were meticulously catalogued for replay attempts against other targets to try and exploit any shared accounts. When the harness deemed it had enough credentials to try on other assets, it appeared to operate under a constraint internally referred to as a “BRUTEFORCE POLICY”. The policy itself was not laid out or defined in any of the recovered files and was only referenced in log entries that tried to re-use passwords.

Figure 10: Agent authentication attempts constrained by “BRUTEFORCE POLICY”
The “rockyou-style mass spray” highlighted above refers to an old but infamous list of compiled passwords leveraged in password cracking attacks for years. Misconfigurations and exposed credentials became the primary method for initial access, and the harness used them to maintain persistence. Eventually, the harness took it a step further and dropped a PHP webshell for added persistence.

Figure 11: PHP webshell deployed
Finally, the harness pursued the goal of most cyber intrusions, data theft and exfiltration. Once the harness established internal access, it quickly leveraged the foothold to exfiltrate data. A wide range of data was targeted, including source code, private keys, and application databases.

Figure 12: A Sampling of Exfiltrated Data
What Defenses Worked?
Thorough study of the harness logs depicts an alarming picture of a capable, automated, and relentless cyber-attack engine, however the same logs signal that some established, tried-and-true defense-in-depth mitigations limited the damage of this campaign.
First, the target had disciplined patch management processes in place, particularly for the amount of edge and perimeter devices the harness could discover and attack in this operation. This caused every exploit attempt by the harness against known CVEs to fail.
When the harness shifted its focus to misconfigurations and vulnerabilities in custom applications, stolen credentials were obtained and quickly used to attempt pivots via credential reuse attacks on other assets. At this stage, identify management proved an effective defense. There are a multitude of log entries where the harness noted that attempts to re-use stolen credentials did not yield successful logins to other systems. This was mainly due to both limited password re-use by the target, and wide deployment of MFA.

Figure 13: Password Re-use Attempts Unsuccessful
Finally, network segmentation isolated the systems that were successfully exploited on the edge, limiting the attack’s spread into internal hosts. Harness log entries below show that attempts to pivot further pivot from the DMZ servers were blocked.

Figure 14: DMZ Network Segmentation Blunted Some Pivots
It is encouraging that long-standing security controls were in place and limited the impact of the agent-driven breach on the victim.
Conclusion
This investigation documents an automated, multi-agent attack harness capable of reconnaissance, exploit selection, victim-specific tooling, and coordinated activity across agents. Its exploit library spanned around 270 packages and targeted CVEs across a multitude of vendor products, disclosed from 2019 all the way through 2026.
This incident joins other high-profile AI security events, including the OpenAI HuggingFace incident and the discovery of Artex, a fully autonomous AI pentesting framework. What sets this investigation apart is that the recovered memory files document an attack as it progressed, revealing how the agents coordinated and adapted to the defenses they encountered.
The harness’s own logs were reviewed and forensically useful in this case, and show that foundational defenses limited the blast radius of this attack:
Patch management and service hardening caused every exploit attempt against known CVEs to fail.
Identity management blocked most reuse of stolen credentials across other assets.
Network segmentation constrained further lateral movement into internal servers.
However, the harness still gained access through misconfigurations and resulted in the exfiltration of source code, private keys, and database information. The incident shows that agent-driven attacks systematically probe for gaps in hardening that remain, and their speed of operation compress the time defenders have to move from initial alert to containment.
Established security practices remain effective against AI-enabled attacks and applying them comprehensively is more important than ever as threat actors mature and refine their playbooks. Defenders should strengthen their security program with:
Rigorous configuration management and patching
Strong credential hygiene
Monitoring tuned to detect high-tempo, systematic activity
Indicators of Compromise (IoCs)
IOC | Type | Notes |
|---|---|---|
2.28.27.38 | IPv4 | operator |
192.99.247.7 | IPv4 | attack-hand1 |
151.158.1.87 | IPv4 | attack-hand2 |