Defenders Use AI to Deploy Adaptive Honeypots for Bots

Security teams are deploying generative AI to create adaptive honeypots that mimic Linux shells and Internet of Things devices from simple text prompts. The systems accept network connections, interact with connecting programs and record attacker activity to catch and analyze automated attack bots.

Engineers assemble the systems from three parts: a TCP listener that accepts connections, a simulated vulnerability that grants access when triggered, and an AI component that generates responses. The listener forwards incoming data to a handler that enforces an access step such as a hardcoded username and password, port knocking, or activation only when a specific exploit pattern appears. After the simulated vulnerability is triggered, the handler passes attacker commands to a generative model and relays the model’s output back to the intruder.

Operators provide a system prompt that directs the AI to impersonate a target environment. Examples include a Linux bash shell modeled as a junior Python developer’s workstation and a BusyBox-based smart fridge with files and configuration consistent with that device. Prompts instruct the model to return only standard output and standard error; low sampling temperature settings are used to keep responses factual and consistent. The AI component requires access to a model API and a valid API key.

Security teams report that the approach scales faster than installing and maintaining traditional honeypot software. By changing a text prompt, defenders can generate many simulated endpoints-developer laptops, routers, cameras and other IoT devices-and deploy them across a network. The deployments are intended mainly to interact with automated tools and AI-driven bots, which often prioritize speed and execution over subtle, human-like judgment.

Researchers report that AI-driven attack agents depend on context and do not possess awareness. Automated agents can be sidetracked by environments that are plausible but not genuine, and may reveal command sequences, payloads and orchestration logic while probing. Some researchers refer to these controlled setups as a ‘hall of mirrors’ where teams can observe attacker behavior in real time.

The method preserves long-standing honeypot goals of gathering data and studying malicious actions while reducing the time needed to create traps. The technique has limits: skilled human adversaries can detect inconsistencies in a simulated environment, and the approach depends on reliable model access and careful prompt design to hold probes long enough to collect useful telemetry.

Honeypots have been used for decades to gather intelligence on malicious activity. Generative AI reduces the configuration effort and allows defenders to expand the number and variety of deceptive endpoints to study fast-moving automated attacks.