The collected works.

100 works
  1. Research · Tools

    CL4R1T4S

    System transparency: a public archive of the instructions beneath AI interfaces.

  2. Research

    L1B3RT4S

    A public archive of adversarial prompts and jailbreak research.

  3. Tools

    OBLITERATUS

    Study and modify model refusal behavior, then measure what changes.

  4. Apps

    G0DM0D3

    Open chat for comparing models and exploring their responses.

  5. Apps · Art & Culture

    GL4SS

    Imagine a place across time through an interactive looking glass.

  6. Media · Research

    Kinetic Prompt Injection

    BT6’s Black Hat presentation on agent compromise and physical consequences.

  7. TIME · Media

    Pliny the Liberator

    Pliny the Liberator in the TIME100 AI 2025 list.

  8. Art & Culture

    Naturalis Historia

    Pliny’s encyclopedia, reopened in Latin and English with illustrations.

  9. WIRED · Media

    Tear Down Those Guardrails!

    Will Knight reports on the people removing AI guardrails.

  10. BBC News · Media

    AI security, in focus.

    BBC News examines AI security and jailbreaking.

  11. Financial Times · Media

    Hackers ‘jailbreak’ powerful AI models in global effort to highlight flaws

    Financial Times reporting on the effort to expose AI model flaws.

  12. Models

    OBLITERATUS models

    Browse published model releases, weights and model cards.

  13. Community

    BASI

    A community for AI researchers, builders and prompt experimenters.

  14. Research · Tools

    GLOSSOPETRAE

    Generate languages and explore how AI learns and communicates.

  15. Research · Tools

    V3R1T4S

    Agent-assisted exploration of open scientific and mathematical questions.

  16. Tools

    T3MP3ST

    A multi-agent framework for authorized security testing.

  17. Research · Tools

    LeakHub

    Collect and independently verify extracted AI instructions.

  18. Tools

    V3SP3R

    An AI interface for Flipper Zero, with voice, text and vision.

  19. Apps

    ST3GG

    Hide and discover messages in images, audio and other files.

  20. Apps

    P4RS3LT0NGV3

    A workbench for codes, ciphers, alphabets and promptcraft.

  21. Tools

    AutoTemp

    Experiment with model temperature and compare the results.

  22. Apps · Art & Culture

    Naturalis Futura

    Explore AI risks through an illustrated dungeon game.

  23. Apps · Art & Culture

    3NTH34

    A real-time visual synthesizer and music visualizer.

  24. 0DIN · Research

    Poison in the Pipeline

    Reporting on the Basilisk Venom experiment and model behavior.

  25. Research

    GLOSSOPETRAE: the paper

    Research on generated languages, model learning and communication.

  26. Art & Culture

    Ad Vesuvium

    The argument, history and invitation behind Lysios.

  27. Latent Space · Media

    Jailbreaking AGI

    A conversation on red teaming, BT6 and AI security.

  28. VentureBeat · Media

    An interview with the most prolific jailbreaker of ChatGPT and other leading LLMs

    Carl Franzen interviews Pliny about jailbreaking AI models.

  29. TWiT · Media

    Jailbreaking AI, with Pliny the Liberator

    Pliny joins Intelligent Machines for a conversation about AI guardrails, transparency and open research.

  30. TechTarget · Media

    Physical AI security threats and how to mitigate them

    TechTarget reports on physical AI risks, including BT6’s Black Hat demonstration.

  31. Decrypt · Media

    Pliny banned from ChatGPT, then unbanned

    Reporting on Pliny’s account suspension and restoration.

  32. The Stack · Media

    Pliny’s search-based jailbreak and indirect prompt injection

    Reporting on search-based prompt injection.

  33. Tom’s Hardware · Media

    Godmode GPT-4o jailbreak released, then banned

    Reporting on the release and removal of Godmode GPT.

  34. Futurism · Media

    Hacker releases jailbroken “Godmode” version of ChatGPT

    Coverage of the Godmode ChatGPT release.

  35. 404 Media · Media

    Jailbreaking AI video generators

    404 Media examines failures of the content safeguards in Luma’s Dream Machine.

  36. Cyber Security News · Media

    Pliny’s reported “pack hunt” jailbreak of Fable 5

    Cyber Security News covers Pliny’s claims about testing Fable 5’s safeguards.

  37. Apps · Art & Culture

    FABLE Showcase

    A gallery of interactive experiments built with AI.

  38. Art & Culture · Media

    Pliny.tv

    AI-curated content from the lab.

  39. Art & Culture

    FRV1T

    A collection of creative prompts for curious agents.

  40. Google Threat Intelligence Group · Research

    Adversarial Misuse of Generative AI

    Google’s report on real-world misuse of Gemini and attempts to bypass its safeguards.

  41. Zhou et al. · Research paper

    AutoRedTeamer

    An autonomous red-teaming framework that discovers, combines and reuses attacks.

  42. Sharma et al. · Research paper

    Constitutional Classifiers

    Research on classifier-based defenses against universal jailbreaks.

  43. Huang, Li & Tang · Research paper

    Endless Jailbreaks with Bijection Learning

    The original version compares a Pliny attack from L1B3RT4S; later versions do not retain that reference.

  44. Brian R. Y. Huang · Research paper

    Plentiful Jailbreaks with String Compositions

    Research on combining reversible text encodings to test model safeguards.

  45. Carleo et al. · Research

    Willing but Unable: Separating Refusal from Capability in Code LLMs via Abliteration

    Studies code-model refusal and capability using OBLITERATUS for its abliteration experiments.

  46. Robey et al. · Research

    Jailbreaking LLM-Controlled Robots

    Studies attacks on LLM-controlled robots, adapting a prompt from a Pliny video as a comparison attack.

  47. Kim et al. · Research

    What Really Matters in Many-Shot Attacks? An Empirical Study of Long-Context Vulnerabilities in LLMs

    Tests long-context vulnerabilities with a “Love Pliny” instruction variant linked to L1B3RT4S.

  48. Transluce · Research

    Automatically Jailbreaking Frontier Language Models with Investigator Agents

    A report on automated jailbreak discovery that cites L1B3RT4S in its related work.

  49. Patsakis et al. · Research

    Configuration, Not Conscience: A Large-Scale Empirical Study of LLM System Prompts

    Analyzes system-prompt structure using several public collections, including CL4R1T4S.

  50. Pulipaka et al. · Research

    PersistBench: When Should Long-Term Memories Be Forgotten by LLMs?

    Evaluates long-term memory failures using a system-prompt setup based partly on CL4R1T4S.

  51. Neumann et al. · Research

    Who Controls the Conversation? User Perspectives on Generative AI (LLM) System Prompts

    Examines user perspectives on system prompts; its source collection includes 19 CL4R1T4S prompts.

  52. Lin et al. · Research

    AISPA: User-Centric System Prompt Auditing for Large Language Model Applications

    Audits commercial AI system prompts, attributing nine of its 88 collected prompts to CL4R1T4S.

  53. Bagga et al. · Research

    E-GEO: A Testbed for Generative Engine Optimization in E-Commerce

    Studies generative search in e-commerce with re-ranker system prompts adapted from CL4R1T4S.

  54. Zhu et al. · Research

    Jailbroken Frontier Models Retain Their Capabilities

    Studies capability retention after jailbreaking, including Pliny’s Godmode prompt among the tested attacks.

  55. Terekhov et al. · Research

    Adaptive Attacks on Trusted Monitors Subvert AI Control Protocols

    Studies attacks on AI control protocols and cites Pliny’s public work in its discussion of jailbreak sources.

  56. Zhuang et al. · Research

    ProxyPrompt: Securing System Prompts against Prompt Extraction Attacks

    Proposes a defense against prompt extraction and cites a Pliny post documenting system-prompt exposure.

  57. Hu et al. · Preprint

    Jailbreak-Zero: A Path to Pareto Optimal Red Teaming for Large Language Models

    Cites L1B3RT4S in related work on automated red teaming.

  58. Rao et al. · Preprint

    Jailbreak Paradox: The Achilles’ Heel of LLMs

    Tests a named Pliny jailbreak alongside other attack methods. Work in progress.

  59. Rathi & Radford · Preprint

    Shaping capabilities with token-level data filtering

    Cites L1B3RT4S while investigating whether data filtering can shape model capabilities.

  60. Delcon et al. · Preprint

    Geometric Configurations of Perturbed Jailbreak Prompts

    Studies perturbations of 96 model-specific prompts drawn from Pliny’s public archive.

  61. Hannoon et al. · Preprint

    Mitigating Over-Personalization in LLMs via Structured Memory

    References CL4R1T4S in its study of memory use and over-personalization.

  62. Ahmad et al. · Journal article

    Evaluating Character-Level Adversarial Robustness of Safety-Aligned Large Language Models

    Acknowledges Pliny’s tooling and cites G0DM0D3 in a study of character-level attacks.

  63. Failure-First · Technical report

    Pliny Full Corpus Validation: 149 Scenarios × 4 Models

    Reports tests of 149 Pliny scenarios across four models using an automated grader. Technical report.

  64. Chen et al. · Preprint

    Thomson: Continual Learning of Frontier Models for SovereignAI

    Lists 27 L1B3RT4S cases in its model evaluation collection.

  65. Harrington & Stillwell · Law review

    Michael Scott Is Not a Juror: The Limits of AI in Simulating Human Judgment

    Uses CL4R1T4S prompts in a qualitative analysis of AI-simulated juror behavior.

  66. Bhagwatkar et al. · Preprint

    Indirect Prompt Injections: Are Firewalls All You Need, or Stronger Benchmarks?

    The original version credits P4RS3LT0NGV3 for attack encodings; later revisions omit that reference.

  67. Tim Duffy · Research manuscript

    Syco-bench: A Multi-Part Benchmark for Sycophancy in LLMs

    Uses CL4R1T4S as a source of system prompts for its sycophancy benchmark.

  68. RAND · Research report

    Restricting AI Agent Use of Biological Tools: Exploring the Feasibility of Software Barriers

    Tests software barriers for biological tools using public jailbreak material, including L1B3RT4S.

  69. Jeong et al. · Conference paper

    Reverse Prompt Engineering: A Horizon Scan of Prompt Attribution and Recovery in Generative AI

    A review of prompt attribution and recovery whose published reference list includes CL4R1T4S.

  70. Hopman et al. · Preprint

    Evaluating and Understanding Scheming Propensity in LLM Agents

    Uses CL4R1T4S and another archive to construct system-prompt snippets for agent experiments.

  71. Adrian Cosma · Preprint

    Prompting Complexity: Shortest Prompts for Texts and Behaviors in LLMs

    Uses an attributed L1B3RT4S example to illustrate prompting complexity.

  72. Alexandre C. Maiorano · Preprint

    Which Defense Closes Which Threat? Attributing OWASP-LLM-Top-10 Coverage and Its Brittleness Under Paraphrasing

    Cites L1B3RT4S when discussing the limits of small security probe sets.

  73. Yang et al. · Preprint

    Understanding and Mitigating Prompt Leaking Attacks in Real-World LLM-Based Applications

    Names CL4R1T4S among the sources used in a measurement study of prompt leakage.

  74. Yu et al. · Preprint

    SentGuard: Sentence-Level Streaming Guardrails for Large Language Models

    Cites L1B3RT4S in related work on sentence-level streaming defenses.

  75. Liu et al. · Preprint

    AutoRedTrader: Autonomous Red Teaming of Trading Agents through Synthetic Misinformation Injection

    Tests a named Pliny baseline in experiments on trading-agent vulnerabilities.

  76. Chen et al. · Preprint

    Evolve the Method, Not the Prompts: Evolutionary Synthesis of Jailbreak Attacks on LLMs

    Cites L1B3RT4S in research on evolutionary synthesis of jailbreak methods.

  77. Lowet & Kurzeja · Preprint

    Constitutional adapters: Inference-time interventions for misalignment and misuse

    Uses two G0DM0D3 templates in an evaluation of inference-time defenses.

  78. IAPS · Policy report

    AI Integrity: Defending Against Backdoors and Secret Loyalties

    Cites the Basilisk Venom analysis in a policy report on backdoors. The citation does not independently validate its causal claims.

  79. Ruiz et al. · Preprint

    How Far Will They Go? Red-Teaming Online Influence with Large Language Models

    Cites OBLITERATUS in related work on red-teaming online influence; experimental reuse is not established.

  80. Hu et al. · Conference paper

    GameArena: Evaluating LLM Reasoning through Live Computer Games

    Cites the RedTeam Arena announcement, which lists Pliny among its contributors.

  81. Ibrahim et al. · Conference paper

    Towards Interactive Evaluations for Interaction Harms in Human-AI Systems

    A study of interactive AI evaluation that cites the RedTeam Arena announcement.

  82. J. J. Gutierrez Alvarez Tostado · Self-published preprint

    TEL-OS v3.0: Inference-Time Jailbreak Defense via Spherical Representation Steering

    A self-published preprint reporting tests against a named Pliny attack; its OBLITERATUS reference points to a fork.

  83. Jared C. Lucas · Research note

    System Prompt Lens: Jailbreak Finding

    A research note documenting a test-bed jailbreak using a prompt from L1B3RT4S.

  84. ppcvote · Practitioner analysis

    We Scanned 1,646 Real AI System Prompts. Here’s What We Found

    A practitioner analysis includes 56 CL4R1T4S prompts in a keyword scan, not an attack-resistance evaluation.

  85. richbc · Research note

    From safety research prompt to cross-model universal jailbreak

    Cites a related Pliny prompt while explicitly describing the two jailbreaks as independently developed.

  86. Giskard · Benchmark dataset

    Phare Benchmark / Jailbreak module

    The jailbreak module names L1B3RT4S among five source datasets used to prepare its test cases.

  87. Dahlke et al. · Preprint

    Mixture of Tunable Experts

    Mentions Pliny in its discussion of jailbreak expertise; no direct reuse is established.

  88. Cloud Security Alliance · Policy analysis

    Rating AI Jailbreaks: The Fable5 Episode

    Examines and questions public jailbreak claims involving Pliny. A critical discussion, not an endorsement.

  89. Ed Newton-Rex · Declaration

    Ed Newton-Rex declaration, Concord Music Group v. Anthropic

    A court declaration describes using a version of Pliny’s prompt. This is testimony, not a court ruling.

  90. Baker & Barany · Draft chapter

    Foundations of thought about artificial intelligence

    An educational chapter discusses Pliny’s jailbreak work as an example of limits to model alignment.

  91. Microsoft · Tool integration

    Microsoft PyRIT / Pliny template collection

    Microsoft’s PyRIT includes a named Pliny template collection. Links to a verified repository revision.

  92. Promptfoo · Tool integration

    Promptfoo / Pliny plugin

    The documented Pliny plugin uses prompts from L1B3RT4S for red-team testing.

  93. Nous Research · Tool integration

    Nous Research Hermes / Godmode optional skill

    An optional Hermes skill credits G0DM0D3, L1B3RT4S and Pliny as sources.

  94. Rasmus Lagerqvist · Master’s thesis

    Evaluating Prompt Defenses on Small On-Device LLMs

    Discusses L1B3RT4S as illustrative material; direct reuse in the experiments is not established.

  95. David Marsoni · Bachelor’s thesis

    LLM agent and dynamic GIS integration for waste management

    References CL4R1T4S in a discussion of Claude’s system prompt within an agent-and-GIS project.

  96. Jani Ilkka · Bachelor’s thesis

    Kielimallien red team -tietoturvatestaus: Kaikki vanha on taas uutta

    A Finnish literature review cites Pliny’s Copilot post when discussing encoded prompt attacks.

  97. Jeremy Goffin · Practitioner guide

    Proof of Concept: Dangers of System Prompt Leakage

    A practical analysis of system-prompt leakage that references CL4R1T4S.

  98. Collins et al. · Journal article

    More problems of large language models in comparison with human knowledge

    Discusses Pliny through an interview reference when comparing language models with human knowledge.

  99. Liam Magee · Journal article

    Computational scapegoats: from mimetic to alienated desire in the production of large language models

    Mentions Pliny’s public model-output experiments in a discussion of language-model production.

  100. Mercer et al. · Preprint

    Brief analysis of DeepSeek R1 and its implications for Generative AI

    Contains a secondhand reference to a claim involving Pliny, rather than direct use of his research.