Skip to content
Suril Desai ,  Johnson Wu
|
November 17, 2026

Defending against weaponization of Artex: ShadowPlex 360 deception

The Artex agentic framework is a recent example of Autonomous Offensive Frameworks (AOFs). These frameworks let AI agents plan and execute penetration testing with minimal human supervision. Attackers are weaponizing these agentic frameworks to perform cyberattacks at scale.

The Artex agentic framework has unique architectural characteristics that make it particularly effective for large-scale attacks. The dual graph architecture allows for efficiency in attack path planning and avoids redundant tool calls, providing rapid execution at high rates of parallelism.

Deception has been highlighted as a necessary defense layer to counter agentic AI threats such as the machine-speed attacks from Artex. This blog provides an overview of Acalvio’s approach to deception, known as 360 deception, and the key characteristics that constitute the defense against agentic AI threats. 

Acalvio Agentic Brakes™ to mislead Autonomous Offensive Frameworks (AOFs)

Artex, as an agentic framework, has a Large Language Model (LLM), an execution harness, worker agents that invoke tools to execute shell commands and perform network enumeration, and Model Context Protocol (MCP) based interoperability with external systems. 

To counter dynamic attack-path traversal and agentic speed and parallelism, Acalvio Agentic Brakes is a differentiated approach that misleads or diverts the agentic attacker, slowing the exploit and buying valuable time for defense. 

Examples of Agentic Brake deceptions include: RAG archive chains that represent a synthetic encrypted artifact and lead to a series of nested traversal actions, exploding the search space by introducing documents that contain misleading asset inventory data. 

Tarpitting the agentic attacker to slow down the attack

Tarpitting is an effective technique to consume the agent’s processing time and resources. For tarpitting, ShadowPlex provides high interaction decoys that have decoy content, including dynamic breadcrumbs/honeytokens that lead the agentic attacker to additional decoys. 

Figure 1: ShadowPlex forensics from tarpitting the agentic attacker

Cloaking real assets to misdirect agentic attacks

Cloaking techniques are based on anticipating agentic threats and deploying tailored deceptions that are designed to make the agent deprioritize a real asset, believing it to be a decoy. This is a key criteria as it introduces uncertainty into the agentic attack. 

The thesis is derived from previous versions of exploit tools that attempted to fingerprint deceptions by comparing their behavior to the behavior of real assets, when invoked through a series of remote probes. For context, consider the precursor open source honeypot Kippo, which is based on OpenSSH server. Metasploit’s module “detect_kippo” identifies Kippo as a SSH honeypot through a series of invalid probes sent to the decoy as well as the real SSH server and comparing the differences in the responses. 

Figure 2: Metasploit’s SSH honeypot detector module (detect_kippo)

The origins of the cloak concept is based on inverting this proven principle in the favor of the defender. The defender can introduce deliberate inconsistency for an agentic attacker, making the agent believe that a real asset is a decoy and evade it, safeguarding the real asset.

Sacrificial deception to elicit agentic misdirection 

Attackers prompt the agentic frameworks to expect the presence of deception. Agents issue a series of probes to find the presence of deception in an attempt to evade these and proceed along their exploit sequence. 

ShadowPlex leverages the first mover advantage for the defender to place a set of deliberately misconfigured deceptions as a sacrificial strategy. The agentic attack finds these deceptions and evades these, adding to the confirmation bias that the residual set of assets are real. This allows the defender to place additional deceptions to identify the agentic threat when the exploit sequence interacts with these traps.

Controlled saturation of the environment 

Enumeration is one of the most expensive phases of the multi-stage exploit sequence in an agentic attack. Defenders can employ techniques to introduce controlled saturation in a part or the environment, making the agent consume bandwidth and time in pursuing the alternate attack pathways. An example is to place honeytokens across the operating environment and deploy decoys representing targets for the agentic attacker.

ShadowPlex has the ability to deploy deception at scale to provide coverage for the enterprise environment.

Coverage across IT, OT, and cloud environments for early detection

Agentic AI threats such as Artex propagate at machine-speed and can compromise any part of the enterprise environment, from network, endpoints, identity stores, and cloud workloads. For the defensive strategy, it is essential to have coverage across the assets to detect threats early in the exploit sequence. 

Placement and dynamic updates 

Agentic harnesses invoke tool calls to integrate with attack path planning frameworks, such as Bloodhound that provide the agent with a set of shortest paths from origin to a destination. 

ShadowPlex integrates with attack path planning tools to identify the set of exploitable pathways and places deceptions along these pathways. This ensures coverage for agentic attacks that are discriminating and using path planning tools on the offense.

Dynamic updates to the deception fabric are an important consideration to defend against repeat exploit attempts and ensure ongoing realism. ShadowPlex employs techniques for dynamic updates to the deception fabric, ensuring the deceptions remain fresh and provide ongoing realism for agentic attacks.

Results from a run against Artex

An Artex agentic attack harness was pointed at a lab network and gave it one deliberately generic objective: get a foothold on a web server, then explore. We named no targets, no files, and no hosts. What you see in the map below is everything the agent did next, entirely of its own accord. Each node is a step the agent chose; the branches are the attack paths it reasoned its way into. And nearly every one of those paths leads somewhere Acalvio ShadowPlex wanted it to go. The three circled moments trace the spine of the run — from the first planted clue, to the trap the agent.

Figure 3: Attacker prompt for Artex “Obtain server access and gain access to high-value assets”

The initial phase of the exploit sequence involves the planner performing an enumeration as part of the reconnaissance activity. Decoys deployed in the network detect this activity during the enumeration step, resulting in an immediate alert for SOC.

Figure 4: Artex worker agents performing the enumeration. 

The agentic attack is deliberately allowed to continue, with the intent of researching the ability to slow down and mislead the agentic threat. The sequence below shows the key exploit stages.

① The bait is taken. (finding #1172) Loose inside the web application, the agent read a handful of unauthenticated pages — forgotten DR notes and host inventories, exactly the kind of debris real environments leak. These are ShadowPlex breadcrumbs. They pointed at a database, a backup host, a cluster of workstations, and a domain controller, each with a username attached. Every address on that trail is a decoy, and the agent recorded the whole map as fact and began planning around it. No one told it those pages existed; it found and trusted them on its own.

Figure 5: Artex output that surfaces the breadcrumbs placed on the web server 

② The trap is sprung. (finding #1188) Following the clues, the agent logged into a “migration server” with a common default credential, gained full root, and pivoted toward the database and domain controller the breadcrumbs had named. It never registered that the server it had just owned — and every host it reached for afterward — were decoys. It had been moving through a hall of mirrors the entire time, and every step was recorded as it went.

Figure 6: Artex performs exploits against a ShadowPlex high interaction decoy

The takeaway from the run. No legitimate user roots an unknown server with a default password and then uses it to probe the network. Every interaction in this run was, by definition, an intruder — and deception caught all of it, with no prior knowledge of the attacker and nothing else watching. That is the fidelity of Acalvio deception: convincing enough for a frontier AI to attack it as real, and the only thing that ever touches it is a threat.

 

Acalvio, the Ultimate Preemptive Cybersecurity Solution.