Deployment environment
Decoy web services, routes and data associated with applications and APIs.
Research question
Definitions and technical considerations should be checked against the primary source. Atlas records are a selection tagged with this topic; a matching tag does not demonstrate operational outcomes.
Cross-reference
These counts describe tagged Atlas records, not the worldwide number of studies or offerings.
Linked library
Database security suite. Database proxy with field-level encryption, search through encrypted data, SQL injections prevention, intrusion detection, honeypots. Supports client-side and proxy-side ("transparent") encryption. SQL, NoSQL.
The authors propose a dynamic honeypot framework combining a reinforcement-learning deception agent (tabular Q-learning, formalized as an MDP), threat-intelligence fusion via indicators of compromise, and automated containment actions. Its purpose is to adapt the honeypot's exposed persona and guide deception decisions, evaluated in an isolated virtualized testbed.
Pacheco and Staino paper in the SACS 2026 proceedings describing the reproducible educational design of Cyber Deception Playground.
This paper proposes seven prompt-injection detection techniques that adapt mechanisms from disciplines outside LLM security, including forensic linguistics, materials-fatigue analysis, deception technology, bioinformatics sequence alignment, mechanism design, spectral analysis, and taint tracking. According to the authors, four of the seven techniques were implemented in the prompt-shield v0.7.3 tool, including honeypot tool definitions, and evaluated through ablations across nine benchmarks totaling approximately 10,300 samples.
This work studies credential exfiltration by LLM agents via indirect prompt injection, proposing three complementary defenses according to the authors: activation probes for pre-output detection, honeytokens calibrated via split conformal prediction, and a cumulative leakage budget tracked across multi-turn conversations. The authors report preliminary results from controlled experiments on open-weight models and a small in-house multi-turn synthetic suite.
Open Docker Compose lab with vulnerable services, configurable deception levels, monitoring and an attacker environment for controlled training and research.