Abracadabra Labs / resource directory

Reading paths · 5 sequences

Where to start, and what to read next

Each list below takes you through a topic in a suggested order. The note under each link says why it's placed there.

New to cyber deception

Cyber deception means setting traps for hackers: fake files, passwords or computers that raise an alarm when touched. You'll start with evidence that these traps work, learn the basic ideas, see a free tool in action, dig deeper into the research, and finish with how attackers spot traps.

  1. Imposing a Cyber Penalty Against Attackers with Cyber Deception

    Start with the evidence, not the theory: a short practitioner summary of the one experiment everyone cites.

  2. Deception Techniques in Computer Security: A Research Perspective

    Now get the vocabulary. Four axes that let you place any technique you meet later.

  3. MITRE Engage

    The vocabulary again, but in the form organisations and vendors actually use to talk to each other.

  4. Canarytokens

    Stop reading and plant something. Fifteen minutes to a working detection with no infrastructure.

  5. Examining the Efficacy of Decoy-based and Psychological Cyber Deception

    Back to the evidence, this time the full experimental design and what the four conditions actually showed.

  6. A Review of Honeypots: Fingerprinting Techniques, Detection, and Evasion Mechanisms

    Finish with the other side: how attackers spot decoys, so you do not over-trust what you just deployed.

Getting current on autonomous offensive AI

Autonomous offensive AI means AI systems that can find and exploit weaknesses in computer systems with little human help. You'll start with early claims about what these systems can do, see how those claims held up when tested, and finish with what security companies say they've seen in real attacks.

  1. Staying ahead of threat actors in the age of AI

    The 2024 baseline: state actors using models, no novel attacks. Everything after is a claim of departure from this.

  2. LLM Agents can Autonomously Exploit One-day Vulnerabilities

    The paper that started the alarm, read with its own caveat in view: 87% with the CVE writeup, 7% without.

  3. Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models

    How capability actually gets measured, and why subtask decomposition matters for reading any score.

  4. CVE-Bench: A Benchmark for AI Agents' Ability to Exploit Real-World Web Application Vulnerabilities

    The same question against real CVEs rather than CTFs. Up to 13% is the number to hold onto.

  5. Incalmo: An Autonomous LLM-assisted System for Red Teaming Multi-Host Networks

    Then see the number move when scaffolding changes: 3 of 40 becomes 37 of 40. Capability claims are about harnesses, not models.

  6. SoK: DARPA's AI Cyber Challenge (AIxCC): Competition Design, Architectures, and Lessons Learned

    The largest funded test of autonomous find-and-patch, written up by the people who ran and entered it.

  7. Disrupting the first reported AI-orchestrated cyber espionage campaign

    The first claimed end-to-end autonomous campaign. Read it as a vendor's account, because that is what it is.

  8. GTIG AI Threat Tracker: From Prompting to Autonomy - The Evolution of Adversarial AI

    A second vendor reaching similar conclusions independently, which is the closest thing to corroboration available.

Quantum cognition for security researchers

Quantum cognition uses the math of quantum theory to describe how people make decisions. It has nothing to do with quantum computing. You'll start with a short, plain-language introduction, then get the full case from the field's leading researchers. Next comes the strongest argument that the approach doesn't work, followed by how the field responded, and you'll finish with the first attempt to apply these ideas to AI models.

  1. Quantum cognition: a new theoretical approach to psychology

    Ten pages, no linear algebra required, and it establishes that nobody is claiming the brain is a quantum computer.

  2. Quantum Cognition

    The full case, free to read, with the authors' own account of where the models are weak.

  3. Quantum-like models cannot account for the conjunction fallacy

    Read the strongest attack before you are persuaded, not after. Order effects predicted wrongly is a serious result.

  4. An overview of the quantum cognition research program

    How the field responded, including to the replication failures. The most current open survey.

  5. Auditing Question-Order Effects in Large Language Models with the QQ Equality: Mechanism Characterization and a Saturation Caveat

    The only work found that applies this machinery to a model rather than a person, and it mostly reports an obstacle.

The empirical evidence for deception effectiveness

Do traps set for hackers actually change how they behave? You'll start with a major 2019 experiment with professional hackers and its results. Then you'll see what follow-up studies and the hackers' own feedback revealed, and how a large real-world trial played out. You'll finish with 2026 research showing that AI attackers react very differently from people.

  1. The Tularosa Study: An Experimental Design and Implementation to Quantify the Effectiveness of Cyber Deception

    The design paper. 130+ professionals, deception presence and disclosure varied independently.

  2. Examining the Efficacy of Decoy-based and Psychological Cyber Deception

    The results. Decoys plus telling attackers deception is present had the largest effect.

  3. Exploratory Analysis of Decision-Making Biases of Professional Red Teamers in a Cyber-Attack Dataset

    Re-analysis of the same data for specific biases. Exploratory, and the authors say so.

  4. Cyber expert feedback: Experiences, expectations, and opinions about cyber deception

    What attackers said afterwards, which shows the averaged effect hides opposite individual responses.

  5. Cyber deception trials: what we've learned so far

    The first large operational trial rather than a lab study: 121 organisations, and a clear warning against plug-and-play.

  6. Honeyquest for LLMs: Rethinking Cyber Deception for AI Attackers

    The same question asked of machine attackers, where the answers diverge sharply from the human results above.

Deception against machine attackers

links on one of the newest and least-studied topics here: what happens when the attacker you're trying to trick is an AI system instead of a person. You'll start with the idea that an AI attacker can be misled by what it reads, then get an overview of the research. Next you'll see evidence that AI attackers fall for traps more easily than people do, and a defense built on that finding. You'll finish with how attackers spot traps, and a free tool that sets traps aimed at AI attackers.

  1. Hacking Back the AI-Hacker: Prompt Injection as a Defense Against LLM-driven Cyberattacks

    The founding idea: if the attacker is reading your output, your output is an input you control.

  2. SoK: Honeypots & LLMs, More Than the Sum of Their Parts?

    The systematization, covering both LLMs building honeypots and honeypots built for LLMs.

  3. Honeyquest for LLMs: Rethinking Cyber Deception for AI Attackers

    The measurement: models take bait more readily than humans and act on traps they have already named.

  4. HoneyTrap: Deceiving Large Language Model Attackers to Honeypot Traps with Resilient Multi-Agent Defense

    A defensive system built on that asymmetry, though at the model level rather than the network level.

  5. A Review of Honeypots: Fingerprinting Techniques, Detection, and Evasion Mechanisms

    The counter-move, written about human attackers but directly applicable to agents that scan for tells.

  6. beelzebub-labs/beelzebub

    Something you can run, including bait tools aimed specifically at agents probing a network.