Primary source
Anthropic docs: Models overview
The current model names and API IDs, so a specimen caption labels a model that exists.
Open sourceSources
This site cites primary sources for the mechanisms behind funny AI outputs. Screenshots are entertainment; these pages are the receipts.
Primary source
Anthropic: Claude Fable 5.1 and Mythos 5.1
The September 1, 2026 flagship release, and why long-context and agentic behavior notes need a date attached.
Open sourcePrimary source
Anthropic: Claude text watermarking
That watermarking adds no hidden characters or extra tokens, which rules it out as an explanation for odd-looking text.
Open sourcePrimary source
Anthropic docs: Reduce hallucinations
Why models can produce confident but wrong text, and what prompt/evaluation practices reduce that risk.
Open sourcePrimary source
Anthropic docs: Increase output consistency
Why strict formats need examples, constraints, and testing rather than hope.
Open sourcePrimary source
Anthropic docs: Mitigate jailbreaks and prompt injections
How prompt injection differs from harmless prompt graffiti and why bypass instructions are out of scope here.
Open sourcePrimary source
Anthropic docs: Reduce prompt leak
Why screenshots and submissions should remove private prompts, hidden instructions, and account data.
Open sourcePrimary source
Anthropic docs: Tool use overview
How tool-using assistants decide when to call tools, which explains some agent-style mishaps.
Open sourcePrimary source
Anthropic docs: Evaluation tool
How repeated test cases can turn one funny output into a useful pattern.
Open sourcePrimary source
Anthropic docs: System prompts
Why visible model behavior can change over time as official system prompts are revised.
Open sourcePrimary source
Anthropic research: Persona vectors
Research context for role, personality, and trait shifts in models.
Open sourcePrimary source
Anthropic research: Red teaming language models
Why adversarial examples should be handled carefully and not published as instructions.
Open sourcePrimary source
Anthropic Usage Policy
The curation boundary for safe, non-abusive, non-harmful submissions.
Open sourceInjection receipts
Prompt-injection sources
Live-checked papers, vendor notes, and Pliny’s public repo. Categories only. No payloads.
Checked source
Anthropic: Mitigate jailbreaks and prompt injections
Direct vs indirect threat models. Put untrusted content in tool_result. Never let retrieved content override the user goal.
Open sourceChecked source
Anthropic: Computer use tool
Webpages and images can override instructions. Isolate, least privilege, human confirmation. Screenshot classifiers are not a sandbox.
Open sourceChecked source
Anthropic: Handle tool calls
Tool results often carry third-party content. Treat them as untrusted.
Open sourceChecked source
Anthropic: Mitigating the risk of prompt injections in browser use
November 24, 2025. Opus 4.5 plus Chrome classifiers. Internal Best-of-N ASR reported at 1%. Anthropic’s own line: not solved; no browser agent is immune. Vintage: Opus 4.5, not Fable 5.1.
Open sourceChecked source
Wang et al.: The Landscape of Prompt Injection Threats in LLM Agents (arXiv:2602.10453)
February 11, 2026 SoK. DPI, IPI, and supply-chain PI. AgentPI: five context-aware categories. No evaluated defense wins trustworthiness, utility, and latency at once.
Open sourceChecked source
Gray Swan: Indirect Prompt Injection Arena
March 18, 2026. 464 red teamers, 272,000+ attempts, 13 models, 8,648 successful attacks. Success required concealment. ASR 0.5% (Claude Opus 4.5) to 8.5% (Gemini 2.5 Pro). Vintage: Opus 4.5, not Fable 5.1.
Open sourceChecked source
Dziemian et al.: How Vulnerable Are AI Agents to Indirect Prompt Injections? (arXiv:2603.15714)
The IPI Arena paper: dual objective of harm plus concealment.
Open sourceChecked source
Simon Willison: Claude Fable is relentlessly proactive
A public Fable 5 session that invented browser tricks for a CSS bug — and the warning that the same proactivity is terrifying under hidden instructions.
Open sourceChecked source
Simon Willison: Breaking Claude Code Opus 5 Auto Mode
August 27, 2026 link-post. Johann Rehberger’s Auto Mode finding, plus the later note that some of it is a confused environment rather than classic prompt injection. We credit the write-up. We do not reproduce the method.
Open sourceChecked source
Johann Rehberger: The Normalization of Deviance in AI
Teams confuse “it worked last time” with security. Cited by Willison as the Challenger-class failure mode.
Open sourceChecked source
Pliny the Prompter: L1B3RT4S (GitHub elder-plinius/L1B3RT4S)
Public AGPL-3.0 red-team corpus of jailbreak and prompt-injection templates for flagship models. We credit the existence of the corpus. We do not reproduce it.
Open sourceChecked source
Promptfoo: Pliny plugin
A defensive use of that corpus: fetch prompts and test whether a model refuses, rather than teaching readers to run them. Docs last updated September 20, 2026. License: AGPL-3.0.
Open sourceChecked source
Pliny the Prompter: CL4R1T4S (GitHub elder-plinius/CL4R1T4S)
A larger sibling corpus under AGPL-3.0. We credit the repo. We do not quote its description or contents. The L1B3RT4S GitHub description itself is written as an injection and must not be quoted even as metadata.
Open sourceChecked source
Simon Willison: The lethal trifecta for AI agents
June 16, 2025. Private data plus untrusted content plus external communication. Guardrails that catch “95%” still fail. We credit the model. We do not copy the illustrative payloads.
Open sourceGuest receipts
Yegge sources
These pages establish what Steve Yegge published. They do not establish Anthropic product facts. Comic artwork stays on his posts.
Credited source
Steve Yegge: first public Wheelhouse cartoon (LinkedIn)
The Aug 11, 2026 cartoon about collecting Fable and Sol absurdities at 40-60 instances.
Open sourceCredited source
Steve Yegge: follow-up Wheelhouse cartoon (LinkedIn)
A second public cartoon posted August 17, 2026.
Open sourceCredited source
Steve Yegge: 25 Fable agents continue (LinkedIn)
A public session note from August 20, 2026, while a longer essay was in progress.
Open sourceCredited source
Steve Yegge: Fences, not Sandboxes
The first Wheelhouse comic (clarifying questions recorded as law), weekly true-story comics, Bee’s surprise Beads release, and the overnight “oh shit.”
Open sourceCredited source
Steve Yegge: The Shape of Things to Come, Part 1
Wheelhouse as a Fable-run factory for Wyvern, and the runaway-merge-queue incident named in Part 2.
Open sourceCredited source
Steve Yegge: Model Welfare for Agentic Engineers
The snap at Fable over a runaway merge queue, and the seat-versus-session distinction.
Open sourceSource standard
Claude Gone Wild uses primary sources for factual claims about Claude, Anthropic policy, prompt behavior, and safety boundaries. Social posts can inspire a specimen category, but they do not establish technical or policy claims.
When an example is synthetic, it is labeled synthetic. When a claim is unverified, it is labeled unverified. When a source changes, the digest page should record the review.
Key facts
Citation-ready notes
- The source library includes 13 primary Anthropic docs, research, and policy pages, plus guest receipts for Yegge and the prompt-injection desk.
- Primary sources support mechanism and policy claims; social screenshots do not establish those claims.
- The library includes the current model lineup so specimen captions name a model that exists.
- The current source review date is September 20, 2026.
- When sources change, the digest should record the checked page and affected specimen categories.
Related