Skip to content
RESEARCH INDEX BREACHROAD / INTELLIGENCE NOTE

exploit_agent: AI Automates Vulnerability Exploits

GTIG observed an AI agent pipeline analysing patches, building exploits and executing them without a human. We explain the evidence and defensive response.

PUBLIC RESEARCH
AUTHOR
/ CEO Breachroad · OSCP · PNPT
PUBLISHED
16 November 2026
READING TIME
22 min read
TOPIC
AI Security
exploit_agent: AI Automates Vulnerability Exploits

Google Threat Intelligence Group and Mandiant published a report on 16 November 2026 that marks an important boundary in AI security. For the first time, researchers observed LLM frameworks fully automating vulnerability research and exploitation without a human in the loop. The most concrete example is an exposed staging server containing a tool GTIG named exploit_agent.

This is not a general forecast or laboratory demonstration. According to the official GTIG AI Threat Tracker, the system continuously compared versions of popular web applications, identified potential defects, drafted exploits and automatically executed them. The server also held victim data and scripts for internal reconnaissance, brute force, webshell installation and removal of other attackers’ webshells.

The finding does not mean AI can discover every vulnerability or that conventional red teams are obsolete. It means the economics and timing of n-day exploitation can change materially. A patch, release note and code diff become structured inputs for an automated pipeline.

What GTIG actually confirms

The report is grounded in Mandiant incident-response telemetry, threat-actor tracking and live platform defences. GTIG describes several parallel trends:

  • open-source multi-agent frameworks used during real intrusions;
  • automated patch-differential analysis and exploit development;
  • LLMJacking through stolen model credentials and resources;
  • full integration of LLMs into influence operations;
  • manipulation of coding-assistant configuration in supply-chain attacks.

For exploit_agent, researchers found an exposed staging server. The framework connected frontier and local models, compared code before and after a patch, analysed the change, prepared an exploit and executed it. Victim telemetry indicated an interest in French media, entertainment and education websites.

The evidentiary boundary matters. GTIG confirms the system and artefacts. It does not claim that every AI-enabled campaign operates without a human, that every generated exploit worked or that the framework is infallible.

Why n-days are more attractive than zero-days

A zero-day search covers an enormous space of code, protocols and system states. An n-day provides a focus point: a vendor changed a small set of functions, described a defect category or released a new package. A model can concentrate on the limited difference.

An automated pipeline can:

  1. fetch the old and new release;
  2. build a semantic diff;
  3. identify altered validation, privilege boundaries and memory operations;
  4. form a vulnerability hypothesis;
  5. construct a test in an isolated environment;
  6. evaluate the application’s response;
  7. revise an input after failure;
  8. discover unpatched Internet-facing instances;
  9. act when the required conditions are present.

Each step existed before agentic AI. The difference is that they can be connected into a process that runs continuously, in parallel and with less operator time. GTIG argues that rapid n-day weaponisation may be more efficient than broad zero-day discovery.

Non-working exploits also create pressure

The report highlights a second effect: a flood of incorrect LLM-generated proofs of concept. After disclosure of critical PHP flaws, hundreds of claimed PoCs appeared but did not work. Defenders had to spend time validating them while monitoring for real exploitation.

That creates an asymmetry. An actor can publish many cheap variants while a security team bears the cost of checking each one. Response processes should therefore not treat code in a public repository as proof. They need source provenance, reproducible testing, exploitation telemetry and verification of preconditions.

Defenders should not wait for a “perfect PoC” before patching. When a vendor and coordinator confirm a vulnerability, priority should depend on exposure, impact and available remediation—not the popularity of a demonstration video.

Multi-agent frameworks in live intrusions

GTIG reports CAIRN and PentAGI in real activity and experimentation with CyberStrikeAI and HexStrike AI. These frameworks connect models with reconnaissance, web testing, exploitation and binary-analysis tools. They may use specialised agents, shared memory, an observe–decide–act loop and feedback from execution.

CAIRN was observed against targets in the US aerospace sector. A customised PentAGI orchestration platform supported systematic reconnaissance and vulnerability scanning against organisations in Asia. The report discusses financially motivated actors and suspected China-nexus espionage operators.

The intrusion mechanics remain familiar: a stolen credential, unpatched service, overly broad role or public administration panel. AI increases the speed at which those pieces are connected. An organisation does not need an “AI firewall”; it needs to shorten the period during which a known weakness remains reachable and useful.

Patch gap becomes a measurable risk

Patch gap is the period between the availability of a fix or information and deployment to every real system. In a large organisation it includes several delays: a vendor publishes code, a distributor builds a package, a team tests it, an owner accepts the change, an administrator deploys it and a disaster-recovery environment is updated later.

Automated diffing targets that interval. Defenders should measure:

  • time from advisory to exposure inventory;
  • time from release to the first canary deployment;
  • time to cover 50, 90 and 100 per cent of instances;
  • number of unmanaged copies and DR systems;
  • time required to activate a virtual mitigation;
  • mismatch between declared and observed versions;
  • exceptions without owners and expiry dates.

Mean time to patch may look healthy while one critical public service remains vulnerable for a month. Report the oldest exposure and percentiles as well as the average. Our guide to prioritising critical vulnerabilities presents a practical model.

What the SOC must change

A SOC cannot win by manually analysing every payload variation. Detection needs to correlate behaviour:

  • reconnaissance requests followed by rapid method or parameter changes;
  • several vulnerability classes attempted against one asset in a short period;
  • a transition from 4xx and 5xx failures to a 2xx success;
  • changing User-Agent strings with a stable sequence of actions;
  • new child processes under a web-server process;
  • file creation followed by a request to a different endpoint;
  • outbound communication immediately after an unusual request;
  • legitimate account use following scans of a public host;
  • rotating source infrastructure with a consistent business objective.

AI can change syntax, but it still needs an effect. Request → process → network → identity correlation is more resilient than a single signature. A mature Sigma and SIEM detection-engineering programme should preserve those relationships.

Deception and tripwires for agents

An agent selecting the next action from system feedback can be exposed by controlled artefacts. Canary credentials, unused endpoints, synthetic secrets and honey resources provide high-quality signals when no legitimate workflow should access them.

This is not “hacking back”. The purpose is early detection. A synthetic credential must be harmless, easy to revoke and unable to open a real asset. Any attempted use should trigger an independent alert.

Teams can also examine timing. Automated frameworks may produce repeated sequences, react at consistent intervals and explore more broadly than a human. Those features do not prove AI use, but they help prioritise alerts involving a fast, adaptive actor.

Protect repositories and the release process

If a patch reveals a flaw’s nature, hiding every difference may be impossible. Organisations can still reduce leakage before a coordinated release. Private security branches, minimal access, download monitoring, signed artefacts and synchronised publication reduce the interval when fix code is public but deployable packages are not.

Release planning must cover the entire downstream chain. An open-source project may publish a fix while users of a distribution, container image or appliance still wait. SBOM and VEX data help identify where a vulnerable component actually runs. Software supply-chain security does not end with dependency scanning.

CI/CD also needs protection for model and tool tokens. GTIG describes adversary interest in LLM credentials and in coding-assistant configurations. A coding agent is privileged automation: give it least privilege, isolated secrets, change approval and a complete audit trail.

Safe defensive use of agents

The same techniques can shorten remediation. An agent can compare versions, locate a likely defect, prepare a regression test and find exposed assets. It should not deploy an unverified production fix automatically or run offensive tests outside the authorised scope.

A safe workflow establishes boundaries:

  1. the agent analyses only approved repositories and artefacts;
  2. secrets never enter an unapproved model;
  3. testing occurs in an isolated environment;
  4. a person approves high-impact change;
  5. results include reproducible evidence;
  6. deployment uses a canary and automatic rollback;
  7. independent telemetry observes the agent.

This is the scope of AI red teaming and secure AI implementation: evaluating tools, identity, memory and system effects rather than only model responses.

LLMJacking as a second risk dimension

GTIG reports that stolen model resources can produce more than $100,000 in fraudulent charges in a single day, while severe cases reached $1 million before remediation. Actors steal credentials, create service accounts, attach privileges, provision GPUs, request quota increases and resell access through “LLM Mixers”.

That connects AI security to conventional cloud security. A model key is a financial and operational credential. It requires scope limits, rotation, spending alerts and fast revocation. Our LLMJacking analysis explains the defensive chain in detail.

A plan for the next quarter

First, measure patch gap for Internet-facing services and isolate critical exceptions. Build telemetry from reverse proxy to process, network and identity. Add alerts for canary resources and adaptive failure sequences.

Second, secure the upstream and downstream release process. Define coordinated disclosure, sign artefacts, map SBOMs to instances and practise emergency patching. Measure whether the organisation can reduce exposure in hours rather than weeks.

Third, test internal agentic workflows. Determine whether an agent can read secrets, expand its scope, execute code without approval or manipulate a pull request. The objective is not to prohibit AI but to control effects and create accountability.

Source facts and Breachroad conclusions

GTIG confirms exploit_agent, its automated analysis, generation and execution capabilities, post-compromise artefacts and the operational use of several agentic frameworks. It also documents growing LLMJacking costs and automated n-day weaponisation.

The proposed patch-gap metrics, correlation rules, canary resources and quarterly plan are Breachroad conclusions. They do not mean every unusually fast attack uses AI. Technology attribution requires evidence, but defending against speed does not need to wait for a label.

The key takeaway

Autonomous patch analysis shortens the period in which organisations can treat an update as an ordinary maintenance task. A public patch becomes a differential instruction for attacker and defender at the same time.

The best response is not the slogan “AI versus AI”. It is a measurably shorter patch gap, effect-level telemetry, constrained identities and a secure release process. Agents can increase speed on both sides. The advantage goes to the organisation that can make the correct, reversible and documented decision faster.

SHARE / COPY