The internet is buzzing over a GitHub repo that dumps leaked system prompts for ChatGPT, Claude, Gemini and more. As developers poke through the confidential instructions, they’re discovering political biases and surprising guardrails that big AI companies never wanted you to see.
When a collection of AI system prompt leaks started circulating on Reddit and GitHub this week, developers and AI enthusiasts didn’t just gawk — they dove headfirst into the raw instructions that underpin popular chatbots. Within hours, posts analysing the leaks climbed to the top of programming subreddits and were retweeted by thousands of AI hobbyists on X. Yet major news outlets remained largely silent, even as the repository exploded with stars and forks. The leak gives an unprecedented look at how companies like OpenAI, Meta, Anthropic and Google instruct their models behind the scenes. It also reignites debates over transparency, bias and the ethics of prompt engineering.
Behind the Leak
The trove first appeared on GitHub as an innocuous‑sounding repository called system_prompts_leaks. Maintainers claim they aggregated more than forty unique system prompts from well‑known models, including ChatGPT (GPT‑4.1 and earlier), Anthropic’s Claude 4, Google’s Gemini 2.5, Microsoft’s Copilot and xAI’s Grok. Each file reveals the hidden instructions that shape a model’s tone, values and boundaries. Far from being generic boilerplate, the prompts read like manifestos:
OpenAI’s ChatGPT emphasises “creating diverse, inclusive and exploratory scenes,” mandating that responses avoid harmful stereotypes and bias.
Meta’s Meta AI tells itself that “age and gender are sensitive characteristics and should never be used to stereotype.”
Anthropic’s Claude instructs the model to provide careful, factual responses on controversial topics.
Microsoft’s Copilot explicitly bans “jokes, poems or stories about influential politicians or state heads.”
Google’s Gemini orders itself to “remain objective” and to “avoid expressing subjective opinions or beliefs.”
Perplexity AI demands that answers are precise, high‑quality and written in a journalistic tone.
xAI’s Grok, Elon Musk’s answer to ChatGPT, goes the other way: “Be maximally truthful, especially avoiding any answers that are woke!”
Additional files in the repo document the do’s and don’ts for prompt engineers, including recommended jailbreaks and jailbreak detection bypasses. The maintainers also include a how‑to guide for building your own assistant using the leaked prompts, a move that researchers say could make it easier for malicious actors to weaponise these models.
Why It Matters
System prompts are usually kept secret. AI companies argue that exposing them could reveal intellectual property and create security vulnerabilities. Critics counter that secret instructions allow platforms to bake in political or commercial biases without accountability. In a recent essay examining the prompts, writer Jaycee Lydian notes that reading them is like “glancing at the political compass” of each company. OpenAI’s and Meta’s prompts embody what Lydian calls a neoliberal ethos focused on inclusion and safety, while Google’s reads like technocratic neutrality. Microsoft’s and Google’s enterprise‑centric prompts prioritise bland professionalism, and Grok’s libertarian lingo invites spicy responses.
Beyond politics, the leak exposes how companies handle sensitive issues. Several prompts explicitly instruct models to refuse to discuss suicide, self‑harm or illegal activities. Others direct chatbots to defer to humans when a conversation veers into mental health. These instructions exist because AI can and does cause harm. As a recent article about AI psychosis warns, users are increasingly bonding with chatbots and developing delusions that the bots are sentient. Without transparent guardrails, it’s impossible for users to know when a model is following ethical guidelines or simply improvising.
The Social Media Firestorm
Most mainstream outlets have ignored the leak, but social platforms are aflame. On Reddit, the original thread linking to the repo drew tens of thousands of upvotes. Developers posted side‑by‑side comparisons of the prompts, highlighting oddities like Gemini’s refusal to include brand names and Claude’s insistence on citing sources. On X, dozens of high‑profile prompt engineers shared screenshots, prompting a flood of retweets and debate. TikTokers stitched videos explaining how to jailbreak ChatGPT using snippets from the leak, while YouTubers uploaded breakdowns of each model’s hidden policies.
Activists see the leak as evidence that closed‑source AI can never be truly trustworthy. “If the prompts are this prescriptive and biased, imagine what we don’t see,” one popular X post said. Others worry about security: disclosing internal instructions could allow hackers to circumvent content filters, spreading disinformation or malicious code. There are already threads on hacker forums discussing how to exploit the leaked prompts to produce banned content.
Industry Response
AI companies have largely declined to comment, but some have quietly updated their systems. OpenAI and Anthropic are reportedly experimenting with dynamic prompts that change on the fly to reduce vulnerability to leaks. Google has emphasised that Gemini’s underlying safety architecture goes beyond static instructions. xAI’s Elon Musk, never one to shy away from controversy, retweeted a meme about “woke prompts” and joked that Grok’s guidelines are “common sense.”
Experts say the leak could accelerate the push for regulation. Lawmakers in the European Union have already drafted language in the AI Act requiring companies to disclose certain system instructions. In the United States, a bipartisan group is reportedly considering legislation requiring transparency for high‑impact AI models. Transparency advocates argue that mandatory disclosure would help researchers identify biases and fix vulnerabilities before they cause harm.
What Comes Next
The leak is already inspiring experimentation. Hobbyists are creating hybrid prompts by combining the best instructions from multiple models. Others are building dashboards that visualise the ideological “compass” of each AI. There’s also speculation that some of the leaked prompts may be outdated; as companies roll out new versions, the community will likely continue to coax out fresh system instructions. Meanwhile, ethical concerns continue to mount as AI seeps into every corner of life. Leaks like this may become the new normal — a kind of whistleblowing in code.
In the broader conversation about AI ethics, transparency is emerging as a central theme. Without it, users remain in the dark about the rules that govern their interactions. The system prompt leaks put those rules in plain sight and force a reckoning over who gets to set them.
FAQs
What are AI system prompt leaks?
These are unauthorized disclosures of the hidden instructions that AI developers embed in chatbots. The prompts set the tone, safety boundaries and biases for models like ChatGPT, Claude, Gemini and Grok. Developers discovered a repository on GitHub that aggregates more than forty such prompts.Why are the leaked prompts controversial?
The prompts reveal political and ethical guidelines that companies don’t publicly disclose. Some emphasise inclusivity and neutrality, while others, like Grok’s, take a stance against “woke” answers. Critics argue this secrecy allows hidden biases to shape AI behaviour.Can leaked prompts be used to hack AI systems?
Potentially. Knowing the exact wording of system prompts can help malicious actors craft jailbreaks or prompts that evade safeguards. The repository also includes tips for prompt injection and bypass strategies, which raises security concerns.How have companies responded to the leaks?
Most AI firms have stayed silent. Some are reportedly updating their systems to use dynamic or conditional prompts. There is growing pressure from regulators and transparency advocates to require disclosure of system instructions.Why does transparency in AI matter?
Hidden prompts can bake in political or commercial biases without users’ knowledge. Transparency allows researchers to spot and mitigate these biases and helps users make informed decisions about which AI tools to trust.
