Simon Willison
Simon Willison is a software developer and writer who co-created the Django web framework, created the open source data tool Datasette and the LLM command-line tool, and coined the term "prompt injection" in September 2022 [1][2][16]. He has written a blog at simonwillison.net since 2002, and in recent years has written mostly about large language models, coding agents and AI security [1]. He also named the "lethal trifecta" of agent capabilities that make data theft through prompt injection possible, and was an early proponent of "slop" as the term for unwanted AI-generated content [1][20][22].
Before going independent, Willison worked as a developer at the Lawrence Journal-World newspaper in Kansas, at Yahoo, and at The Guardian in London, co-founded the conference directory Lanyrd, and was an engineering director at Eventbrite after it acquired Lanyrd in 2013 [1][3][4][6]. As of November 2026 he serves on the board of directors of the Python Software Foundation, a seat he has held since 2022 [1][28][30].
Early career and Django
Willison was a university student in England in 2003 when he read a job post on Adrian Holovaty's blog for a web developer in Lawrence, Kansas. His university allowed students to spend a year working in industry, and he asked whether the job could become a year-long internship [4]. He spent 2003-2004 at the Lawrence Journal-World, whose web operation was known as World Online [2][3].
According to the Django project's own FAQ, the World Online developers, Holovaty and Willison, dropped PHP for Python in fall 2003 and, while building sites such as Lawrence.com, began to extract a generic web development framework from their code [2]. In Willison's account, both men were tired of maintaining large PHP codebases, and they wrote their own thin layer over mod_python because they doubted it would scale; that layer became Django's request and response objects [3]. He wrote that they pair-programmed core parts of the framework, including the URL resolver and the template language, and that for a long time the code was simply called "the CMS" [3]. Shortly before his internship ended, the team hired Jacob Kaplan-Moss as his replacement, and Holovaty and Kaplan-Moss later persuaded the newspaper's owner to release the framework as open source [3]. World Online open-sourced Django in summer 2005 [2].
After the internship, Willison finished his degree in England and joined Yahoo, working from the company's UK office for a research and development team based in the United States for about a year and a half [4]. He then worked as a consultant, much of it on OpenID, the decentralized sign-in standard, which he backed because he feared a single company would otherwise come to control sign-on for the whole web [4].
The Guardian
Willison later joined The Guardian in London, where he spent two years on data journalism projects [8]. One project he has described in detail was a crowdsourcing site built in Django during the UK parliamentary expenses scandal. When the government released the expense documents, The Guardian asked readers to review random pages and flag anything worth an investigative reporter's attention; readers could enter a postcode to see their own MP's claims. Willison said he built a proof of concept on the first of about five days' notice, and that readers worked through hundreds of thousands of pages within 18 hours of launch [4]. He later said that the problems he saw at The Guardian inspired Datasette [8].
Lanyrd and Eventbrite
While at The Guardian, Willison began building side projects with Natalie Downe, a front-end developer who became his wife [4]. The pair created the first version of Lanyrd, a conference directory website, as a side project while on honeymoon in Casablanca, and launched it at the dConstruct conference in 2010 [5]. Willison's own biography describes Lanyrd as a Y Combinator-funded company that he co-founded in 2010 [1].
On September 3, 2013, Eventbrite announced that it had acquired Lanyrd, together with the Argentina-based Eventioz, for an undisclosed sum. The Next Web reported that Eventbrite's chief executive, Kevin Hartz, cited Lanyrd's understanding of structured data as the reason for the purchase, and that Willison's arrival was notable because Eventbrite's platform was built on Django [6]. Willison joined Eventbrite through the acquisition and was an engineering director there before going independent [1].
JSK fellowship and Datasette
In May 2019 Willison announced that he had been accepted into the John S. Knight (JSK) Journalism Fellowships program at Stanford for the 2019-2020 class [7]. His project was "Improving the impact of investigative stories by expanding the open-source ecosystem of tools that allows journalists to share the underlying data" [7]. Writing when the fellowship began in September 2019, he said his aim was to make the kind of data reporting done by well-funded papers such as The New York Times and The Washington Post more accessible to smaller newsrooms without full-time software engineers [8].
Datasette predates the fellowship. Willison released its first public version on November 13, 2017, describing it as a tool for creating and publishing JSON APIs for SQLite databases; his launch demos served databases built from Creative Commons licensed CSV files published by FiveThirtyEight [9]. The project now describes itself as "an open source multi-tool for exploring and publishing data", aimed at data journalists, museum curators, archivists, local governments, scientists and researchers [10]. Its site lists a wider ecosystem of tools and plugins [10]. Willison also wrote sqlite-utils, a command-line tool and Python library for manipulating SQLite databases that he first released in July 2018 [12]. As of November 2026 the latest stable Datasette release on PyPI is 0.65.6, while development continues on a long series of 1.0 alpha releases (1.0a43 was uploaded on November 15, 2026) [11]. His disclosures state that part of the work on Datasette Cloud is sponsored by Fly.io [1].
LLM command-line tool
Willison published the first release of LLM, a command-line tool and Python library for working with large language models, on PyPI on April 1, 2023 [13]. Version 0.5, released on July 12, 2023, added a plugin system so that plugins could add support for additional models, including models that run on the user's own hardware; the first plugins covered 17 openly licensed models from the GPT4All project, Mosaic's MPT-30B and Google's PaLM 2 [15]. The documentation now describes LLM as a tool for working with models from OpenAI, Anthropic, Google and many others, both through remote APIs and as locally installed models. It stores prompts and responses in SQLite, generates and stores embeddings, extracts structured content from text and images, and lets models execute tools [14]. The latest release as of November 2026 is version 0.38, published on November 17, 2026 [13].
Writing on AI
Willison's blog combines long articles, short link posts and notes, and he also publishes a roughly weekly newsletter and a monthly paid briefing for sponsors [1]. Ars Technica has described him as an "independent AI researcher" [27]. When Claude 3 Opus overtook GPT-4 on the LMSYS Chatbot Arena leaderboard in March 2024, he told the publication it was the first time the best available models came from a vendor other than OpenAI, adding that "we all benefit from a diversity of top vendors in this space" [27]. In June 2025 he gave an invited keynote at the AI Engineer World's Fair in San Francisco, his third talk at the event after October 2023 and June 2024 [26].
Prompt injection
In September 2022 Riley Goodside showed that GPT-3 could be told to ignore its earlier instructions by text placed in its input. On September 12 Willison published "Prompt injection attacks against GPT-3", arguing that this was not a curiosity but a security exploit, and wrote: "I propose that the obvious name for this should be prompt injection" [16]. He chose the name by analogy with SQL injection: applications built on GPT-3 assembled their prompts by joining trusted instructions to untrusted user input, much as vulnerable web applications built SQL queries from strings [16][17]. The same post showed that the technique could also leak an application's hidden prompt [16].
In a March 2024 post Willison argued that prompt injection and jailbreaking are different attacks that people often confuse. In his definitions, prompt injection targets applications built on top of LLMs by concatenating untrusted input with a developer's trusted prompt, while jailbreaking tries to subvert the safety filters built into the models themselves [17]. He considers the risks from prompt injection far more serious, because it attacks the applications built on the models rather than the models themselves [17].
The dual LLM pattern
In April 2023 Willison proposed the "Dual LLM pattern" as a way to build AI assistants that could resist prompt injection [18]. The design pairs a Privileged LLM, which takes instructions only from trusted sources such as the user and has access to tools, with a Quarantined LLM, which handles untrusted content and has no tool access. A controller written in ordinary code passes references to tainted content (variables such as $VAR1) rather than the content itself, so that unfiltered output from the quarantined model never reaches the privileged one [18]. In the same post Willison called the proposal "pretty bad": he expected it to bring much more implementation complexity and a degraded user experience, and it remained open to social engineering attacks that trick users into pasting untrusted content in or private data out [18]. In March 2025 researchers from Google, Google DeepMind and ETH Zurich published "Defeating Prompt Injections by Design", a paper describing a system called CaMeL [19]. In April 2025 Willison added an update to his 2023 post saying that the paper identified a potential flaw in the Dual LLM proposal and described a more developed system that addressed it [18].
The lethal trifecta
In June 2025 Willison described what he called the "lethal trifecta" for AI agents: access to private data, exposure to untrusted content, and the ability to communicate externally in a way that could be used to steal data [20]. If a single agent combines all three, he wrote, an attacker can plant instructions in content the agent reads and trick it into sending the private data to the attacker [20]. He cited reported attacks of this kind against products including Microsoft 365 Copilot, GitHub's official Model Context Protocol server and GitLab's Duo chatbot [20]. He was skeptical of "guardrail" products that claim to catch most attacks, noting that "in web application security 95% is very much a failing grade", and concluded that for end users who combine tools themselves, "The only way to stay safe there is to avoid that lethal trifecta combination entirely" [20].
| Concept | Willison post | Core idea |
|---|---|---|
| Prompt injection | 2022-09-12 | Untrusted text joined to a trusted prompt can override the developer's instructions [16] |
| Dual LLM pattern | 2023-04-25 | Keep untrusted content in a tool-less quarantined model, pass only references to a privileged model [18] |
| Prompt injection vs. jailbreaking | 2024-03-05 | Injection attacks applications; jailbreaking attacks a model's safety filters [17] |
| Slop (endorsing the term) | 2024-05-08 | Name for unreviewed AI-generated content pushed on people who did not ask for it [21] |
| Lethal trifecta | 2025-06-16 | Private data + untrusted content + external communication in one agent enables data theft [20] |
Slop
On May 8, 2024, Willison endorsed a post by the Twitter user @deepfates, who wrote that "slop" was becoming the term for unwanted AI-generated content, in the way "spam" became the term for unwanted email [21]. He wrote that sharing "unreviewed content that has been artificially generated with other people is rude", and that while not all AI-generated content is slop, content that is "mindlessly generated and thrust upon someone who didn't ask for it" is [21]. The Guardian, reporting on the term later that month, called him "one of the early proponents of the term 'slop'" and quoted him saying that "having a name for this is really important, because it gives people a concise way to talk about the problem" [22].
AI-assisted programming
Willison writes extensively about using LLMs to write software. In March 2025, weeks after Andrej Karpathy coined "vibe coding", he argued that the term should keep its original meaning of building software with an LLM without reviewing the code it writes, rather than covering every form of AI-assisted programming [23]. In February 2026 he began publishing "Agentic Engineering Patterns", a guide built as a series of chapters on getting good results from coding agents such as Claude Code and OpenAI Codex [24][25]. He uses "agentic engineering" for professional engineers using agents that can both generate and execute code to amplify their existing expertise, and places it at the opposite end of the scale from vibe coding [24].
Informal model testing
Willison has said he gets little value from published benchmark numbers and has lost some trust in leaderboards, so he relies on his own informal test: asking each new text model to generate an SVG image of a pelican riding a bicycle [26]. He described the test, which "started as a joke", as unreasonably hard for text models, and used its results to illustrate his June 2025 conference keynote on the previous six months of model releases [26].
Python Software Foundation
Willison was elected to the board of directors of the Python Software Foundation (PSF) in 2022 [28]. He was re-elected for a second three-year term in September 2025, alongside Jannis Leidel, with Abigail Dogbe and Sheena O'Connell joining the board for the first time [29][30]. In October 2025 the PSF withdrew its application for a $1.5 million US National Science Foundation grant intended to improve the security of the Python ecosystem, which it had been recommended to receive, after the agreement added a requirement that the grantee affirm it did not operate programs that "advance or promote DEI". Willison wrote that the PSF's legal advisors said the term would apply to all of the foundation's activities, and that he was one of the board members who voted, unanimously, to turn down the funding [31].
Independent work and disclosures
Willison's biography describes him as an independent open source developer [1]. His published disclosures say that GitHub Sponsors support his open source work, that he does one day of work a week for Prime Radiant, Jesse Vincent's applied AI research lab, and that he is a GitHub Star, an unpaid position [1]. GitHub paid him as a participant in its GitHub Accelerator program in 2023, and Mozilla supported his work through its MIECO program in 2023-2024 [1]. He states that he has not accepted payments from LLM vendors, with one exception: OpenAI paid him for his time at a GPT-5 preview event that was used in a video. He also says he is often invited to preview products from organizations including OpenAI, Anthropic, Gemini and Mistral, often under NDA or embargo and often with free API credits [1]. Since February 2026 his blog and newsletter have carried weekly sponsorship banners, which he says have no influence over his editorial content [1].
Timeline
| Date | Event |
|---|---|
| 2002 | Starts his blog at simonwillison.net [1] |
| 2003-2004 | Year-long internship at the Lawrence Journal-World; work with Adrian Holovaty that became Django [3] |
| Summer 2005 | World Online releases Django as open source [2] |
| 2010 | Co-founds Lanyrd with Natalie Downe [1][5] |
| 2013-09-03 | Eventbrite announces its acquisition of Lanyrd [6] |
| 2017-11-13 | First public release of Datasette [9] |
| 2019-09 | Begins JSK Journalism Fellowship at Stanford [8] |
| 2022 | Elected to the Python Software Foundation board [28] |
| 2022-09-12 | Coins "prompt injection" [16] |
| 2023-04-01 | First release of the LLM command-line tool [13] |
| 2023-04-25 | Proposes the Dual LLM pattern [18] |
| 2025-06-16 | Names the "lethal trifecta" [20] |
| 2025-09-16 | Re-elected to the PSF board [29] |
| 2026-02-23 | Begins the Agentic Engineering Patterns guide [24] |
Continue reading
- Prompt injectionPrompt injection is a class of security vulnerabilities in which an attacker crafts malicious input designed to override, subvert, or manipulate the instructions governing a large language model (LLM).Mentioned 13 times above
- Jailbreak (artificial intelligence)A jailbreak in artificial intelligence is a technique that bypasses the safety guardrails, content policies, and alignment constraints built into large language models (LLMs) and other AI systemsMentioned 4 times above
- Vibe codingVibe coding is a software development practice in which a programmer describes their intent in plain natural language and relies on a large language model (LLM) to generate the corresponding source code.Mentioned twice above
- AI AgentsAn AI agent is a software system that selects and performs actions in an environment in pursuit of an objective.Listed in See also
- Model Context ProtocolModel Context Protocol (MCP) is an open protocol for exchanging context and invoking capabilities between applications that use large language models and external programs or data sources.Listed in See also
- Large Language ModelA large language model (LLM) is an artificial intelligence system built on a transformer neural network with billions to trillions of parameters, trained on massive text corpora to predict the next token in a…Mentioned twice above
More to read
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16Willison, Simon. "About Simon Willison". Simon Willison's Weblog. simonwillison.net/about
- ^1 ^2 ^3 ^4 ^5Django Software Foundation. "FAQ: General". Django documentation. docs.djangoproject.com/...general
- ^1 ^2 ^3 ^4 ^5 ^6Willison, Simon. "What is the history of the Django web framework? Why has it been described as 'developed in a newsroom'?". Simon Willison's Weblog, 24 August 2010. simonwillison.net/...what-is-the-history
- ^1 ^2 ^3 ^4 ^5 ^6Willison, Simon. "Happy 20th birthday Django! Here's my talk on Django Origins from Django's 10th". Simon Willison's Weblog, 13 July 2025. simonwillison.net/...django-birthday
- ^1 ^2Creative Bloq Staff. "Lanyrd founders talk responsive design and acquisition". Creative Bloq, 17 February 2014. creativebloq.com/...lanyrd-interview-21410701
- ^1 ^2 ^3Yeung, Ken. "Eventbrite acquires Lanyrd and Eventioz to speed up its global expansion and aid event discovery". The Next Web, 3 September 2013. thenextweb.com/...xpansion-and-aid-event-discovery
- ^1 ^2Willison, Simon. "JSK Journalism Fellowships names Class of 2019-2020 (and I'm in it!)". Simon Willison's Weblog, 1 May 2019. simonwillison.net/...jsk-fellowship
- ^1 ^2 ^3 ^4Willison, Simon. "My JSK Fellowship: Building an open source ecosystem of tools for data journalism". Simon Willison's Weblog, 10 September 2019. simonwillison.net/...jsk-fellowship
- ^1 ^2Willison, Simon. "Datasette: instantly create and publish an API for your SQLite databases". Simon Willison's Weblog, 13 November 2017. simonwillison.net/...datasette
- ^1 ^2Datasette. "Datasette: An open source multi-tool for exploring and publishing data". datasette.io. datasette.io
- ^Python Package Index. "datasette". PyPI. pypi.org/...datasette
- ^Python Package Index. "sqlite-utils". PyPI. pypi.org/...sqlite-utils
- ^1 ^2 ^3Python Package Index. "llm". PyPI. pypi.org/...llm
- ^Willison, Simon. "LLM: A CLI utility and Python library for interacting with Large Language Models". LLM documentation. llm.datasette.io/...stable
- ^Willison, Simon. "My LLM CLI tool now supports self-hosted language models via plugins". Simon Willison's Weblog, 12 July 2023. simonwillison.net/...llm
- ^1 ^2 ^3 ^4 ^5 ^6Willison, Simon. "Prompt injection attacks against GPT-3". Simon Willison's Weblog, 12 September 2022. simonwillison.net/...prompt-injection
- ^1 ^2 ^3 ^4Willison, Simon. "Prompt injection and jailbreaking are not the same thing". Simon Willison's Weblog, 5 March 2024. simonwillison.net/...prompt-injection-jailbreaking
- ^1 ^2 ^3 ^4 ^5 ^6Willison, Simon. "The Dual LLM pattern for building AI assistants that can resist prompt injection". Simon Willison's Weblog, 25 April 2023. simonwillison.net/...dual-llm-pattern
- ^Debenedetti, Edoardo; Shumailov, Ilia; Fan, Tianqi; Hayes, Jamie; Carlini, Nicholas; Fabian, Daniel; Kern, Christoph; Shi, Chongyang; Terzis, Andreas; Tramèr, Florian. "Defeating Prompt Injections by Design". arXiv:2503.18813, March 2025. arxiv.org/...2503.18813
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Willison, Simon. "The lethal trifecta for AI agents: private data, untrusted content, and external communication". Simon Willison's Weblog, 16 June 2025. simonwillison.net/...the-lethal-trifecta
- ^1 ^2 ^3Willison, Simon. "Slop is the new name for unwanted AI-generated content". Simon Willison's Weblog, 8 May 2024. simonwillison.net/...slop
- ^1 ^2Hern, Alex; Milmo, Dan. "Spam, junk ... slop? The latest wave of AI behind the 'zombie internet'". The Guardian, 19 May 2024. theguardian.com/...f-ai-behind-the-zombie-internet
- ^Willison, Simon. "Not all AI-assisted programming is vibe coding (but vibe coding rocks)". Simon Willison's Weblog, 19 March 2025. simonwillison.net/...vibe-coding
- ^1 ^2 ^3Willison, Simon. "Writing about Agentic Engineering Patterns". Simon Willison's Weblog, 23 February 2026. simonwillison.net/...agentic-engineering-patterns
- ^Willison, Simon. "Agentic Engineering Patterns". Simon Willison's Weblog. simonwillison.net/...agentic-engineering-patterns
- ^1 ^2 ^3Willison, Simon. "The last six months in LLMs, illustrated by pelicans on bicycles". Simon Willison's Weblog, 6 June 2025. simonwillison.net/...six-months-in-llms
- ^1 ^2Edwards, Benj. "'The king is dead': Claude 3 surpasses GPT-4 on Chatbot Arena for the first time". Ars Technica, 27 March 2024. arstechnica.com/...hatbot-arena-for-the-first-time
- ^1 ^2 ^3Willison, Simon. "Weeknotes: Joining the board of the Python Software Foundation". Simon Willison's Weblog, 30 July 2022. simonwillison.net/...psf-board
- ^1 ^2Willison, Simon. "Announcing the 2025 PSF Board Election Results!". Simon Willison's Weblog, 16 September 2025. simonwillison.net/...25-psf-board-election-results
- ^1 ^2Willison, Simon. "The 2025 PSF Board Election is Open!". Simon Willison's Weblog, 9 September 2025. simonwillison.net/...psf-board-election
- ^Willison, Simon. "The PSF has withdrawn a $1.5 million proposal to US government grant program". Simon Willison's Weblog, 27 October 2025. simonwillison.net/...psf-withdrawn-proposal
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
1 revision · v1 · 3,278 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check before publication on November 18, 2026: 78 claims checked against 36 cited and independent sources; 0 material and 6 minor corrections applied before publication.
Cite this page: AI Wiki. "Simon Willison." aiwiki.ai, updated 18 Nov 2026, fact-checked 18 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/simon_willison
