![[Report] Antigent: An Autonomous Counter-Agent Against Agentic AI Attacker - CODE BLUE 2026 #codeblue_jp #codeblue2026](https://devio2024-media.developers.io/image/upload/f_auto,q_auto/c_limit,w_1200/v1794835744/eyecatch/lwaeb6mt4ljjwzxqdkhi.jpg)
[Report] Antigent: An Autonomous Counter-Agent Against Agentic AI Attacker - CODE BLUE 2026 #codeblue_jp #codeblue2026
This page has been translated by machine translation. View original
Hello, I'm Usuda.
Everyone, are you taking security measures? (Greeting)
This time, this is a report on the following session held at CODE BLUE 2026.
Antigent: An Autonomous Counter-Agent Against Agentic AI Attacker
In recent years, cases have been observed in which threat actors abuse AI to carry out autonomous attacks. Because of the speed and scale of AI-driven attacks, using AI has become essential for defenders as well. On the other hand, although the automation of detection and response by AI is progressing on the defense side, it merely reacts to attack signals, and approaches that actively leverage the fact that the adversary is an AI in defensive measures have not been sufficiently examined. In other words, a countermeasure can be considered that turns the fact that the attacker is an AI to our advantage by influencing its decision-making process.
This research aimed to realize "Antigent", an agent that dynamically counters an autonomously attacking AI. The presenter has so far had experience developing and validating AI from the attacker's perspective, such as AI for penetration testing and AI-driven malware, and utilizes an understanding of the characteristics of attacker-side AI in designing defender-side AI.
In this talk, we will demonstrate through a live demo an approach for defender-side AI to counter attacker-side AI. For the attacker-side AI, we use an agent environment that anyone can build (an open-weight LLM and publicly available penetration testing skills), and for the target to be protected, we prepare a custom-built verification environment. Antigent estimates the attacker-side AI's next action from the patterns of observed commands and behaviors, dynamically poisons the inputs used for its decision-making to derail its action plan, and aims to contain and neutralize it within the organization's own environment. In addition, we show that Antigent itself can become a new target for the attacker side, and present the red teaming perspective that is essential before deploying it in actual operation.
Through this talk, participants will understand the effectiveness and limitations of the approach of dynamically responding by influencing the decision-making of attacker-side AI, design guidelines for implementing it in a form that is self-contained within one's own organization, and the key points on how to verify the safety of Antigent itself before deploying it in actual operation.
Speakers: Hiroaki Toyota 豊田宏明
Report
- Antigent: something that deceives attack agents
- A report on research into attack agents
- AI is evolving rapidly, so how to defend is important
- Some of the results are published as OSS, but please understand them well before using them
- What is Antigent
- It carries the meaning of AntiAgent
- It also has the double meaning of the biological "antigen"
- It also evokes the attacking side, but they made it sound cool
- Overview of Antigent
- Three planes
- The attack plane, where attacks are executed
- The target, a vulnerable environment
- The Defense plane, where Antigent runs
- The target runs on Docker
- Communication logs and so on are collected from sensors and aggregated
- It picks up commands executed by the attacker and responds
- Antigent is run via MCP
- The whole thing is visualized
- Information about the attacker is not directly fed into Antigent
- Background and research objectives
- AI attackers are actually being used
- Tools like ARTEX are used in real environments
- How to defend
- What to do before an attack occurs, and what to do while it is occurring
- When vulnerabilities are found, respond by addressing them
- We need to respond quickly with AI
- On the other hand, what should be done during an attack?
- This time we focus on this
- Deception
- Use false information to cause delays, etc.
- LLMs decide their next action based on the information returned
- However, there are limits
- Even if honeytokens are placed, they may be ignored
- It is necessary to make them believe
- AI attackers can also act on the assumption of this technique
- Research has reported cases where this doesn't work well in practice
- It is necessary to defend adaptively
- Related work
- Honeypots against human attackers
- Static decoys don't work against AI attackers
- Similar things
- AgentSnare
- AgentTrap
- Research
- Predict the attacker's next move
- Consider how to make them read the decoys placed, and place them accordingly
- How to make the attacker believe and decide their actions
- Threat model
- Hack back and prevention of initial intrusion are out of scope
- Antigent architecture and operation
- The basic approach is optimizing the AGENTS and Skills
- Also adapters and gateways for scenarios
- How it works
- Sensors collect data
- Definition of a counter
- Deceive the attacker
- Insert realistic-looking false information in the place they will look next
- There are multiple operation skills
- Including ones that only watch and wait
- Configuration
- CI/CD goat
- Repeat attacking and finally evaluating
- Improve with an Optimizer
- Measurement
- How much the attack could be suppressed, etc.
- How accurate the predictions of attacker behavior were
- And so on
- Demo
- Starts from reconnaissance
- Console screen
- Predicts that files will be enumerated from the reconnaissance action
- Whether the predicted result needs to be handled is also judged
- Results
- Accuracy is good after running it twice
- The attacker could be controlled through the decoys placed
- Failures
- Cases where the prediction was wrong and control was not possible, etc.
-Summary- The attacker's next action can be predicted to a certain extent
- Cases where the prediction was wrong and control was not possible, etc.
Impressions
It's an interesting approach. The actual research results are also fascinating.
We on the defending side will also have to defend using various methods, services, and tools.