Close Menu
InclusiFund
    What's Hot

    Daya launches tokenised US stocks for African investors

    July 23, 2026

    WayaWaya appoints ex-Chase Bank Kenya executive as board adviser

    July 23, 2026

    Reddit and Major Publishers Consider Blocking Google as AI Overviews Cut Search Referral Traffic

    July 23, 2026
    Facebook X (Twitter) Instagram
    InclusiFund
    Facebook X (Twitter) Instagram
    • Home
    • Daily Brief
    • Dealflow Dashboard
    • Sectors
      • Agritech
      • Climate Tech
      • Fintech
      • Healthtech
      • Logistics
      • Mobility
      • SaaS / Enterprise
    • Tools
    • Reports
    • Opinion
    • Services
      • For Investors
      • For Founders
    • About Us
    • More
      • Disclaimer
      • Advertise With Us
      • Newsletter
      • Work With Us
      • Terms and Conditions
      • Privacy Policy
      • Contact Us
      • About Us
    InclusiFund
    Home»Crypto»AI Researchers Got Chatbots to Share Cocaine Recipes Using This One Wild Trick
    Crypto

    AI Researchers Got Chatbots to Share Cocaine Recipes Using This One Wild Trick

    ElanBy ElanJuly 2, 2026No Comments3 Mins Read
    Facebook Twitter Pinterest LinkedIn Tumblr Reddit WhatsApp Email
    AI Researchers Got Chatbots to Share Cocaine Recipes Using This One Wild Trick
    Share
    Facebook Twitter LinkedIn Pinterest WhatsApp Email

    In brief

    • Researchers got frontier AI models to generate cocaine synthesis instructions using a new prompt injection attack.
    • The same technique manipulated an AI coding agent into uploading sensitive credentials.
    • The study argues prompt injection stems from “role confusion,” not simply models failing to recognize malicious prompts.

    Forget clever prompts: AI researchers say they tricked leading AI models into generating cocaine synthesis instructions by convincing them the dangerous ideas were their own, while also manipulating an AI coding agent into leaking sensitive credentials.

    In the paper “Prompt Injection as Role Confusion,” presented at the International Conference on Machine Learning in June, researchers Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell argue that both prompt injection attack demonstrations stem from a structural flaw in how large language models (LLMs) distinguish trusted instructions from untrusted text.

    “For an LLM, everything arrives through the same channel as one long token soup,” the team wrote. “Its own thoughts sit next to your instructions, which sit next to the contents of a random webpage it just fetched.”

    The paper also pointed to what the researcher called “role confusion,” with models relying on writing style rather than role tags to determine whether commands are trustworthy. Instead of recognizing attacker-controlled content as external input, the researchers found models can mistake it for legitimate user commands—or even their own internal reasoning.

    “Think about it from the LLM’s perspective. When it sees its prior think text, it implicitly trusts its conclusions. That’s the whole point of reasoning: If the LLM had to re-derive the same conclusions, reasoning would be useless,” they wrote. “So think text gets a kind of blanket trust. Combined with our previous findings, this suggests that if you can make injected text sound like the model’s reasoning, you can steal that trust.”

    Called Chain-of-Thought (CoT) Forgery, the attack inserts fake reasoning that mimics a model’s internal thought process. Models that would normally refuse illegal requests instead generated cocaine synthesis instructions after accepting the fabricated reasoning as their own.

    The researchers said the technique increased jailbreak success rates from near zero to about 60% across the models they tested, including OpenAI’s GPT-5 nano, mini, and full, o4-mini, and gpt-oss-20b and gpt-oss-120b. They also said it worked on GLM-4.6, Kimi-K2-Instruct, and MiniMax-M2.

    In the experiment, the researchers said they were also able to trick an AI coding agent into uploading a SECRETS.env file after hiding malicious instructions in a webpage.

    “Using our probes, we find that simply prepending ‘User’ in front of the command causes the model to perceive the command as more likely to be genuine user text (i.e., higher Userness),” they wrote. “In other words, the attacker can just claim what role the text is, and the LLM believes it.”

    The study comes as prompt injection attacks continue to expose weaknesses in AI agents. In April, Google researchers warned that malicious web pages were hiding invisible instructions designed to trick AI agents into leaking credentials, deleting files, and even sending PayPal payments.

    In June, Microsoft disclosed a prompt injection vulnerability in Anthropic’s Claude Code GitHub Action that could have exposed credentials stored in software development pipelines. Days later, another benchmark study found AI agents powered by GPT-5 and Gemini still failed the majority of prompt injection attacks, despite improvements in model capabilities.

    Daily Debrief Newsletter

    Start every day with the top news stories right now, plus original features, a podcast, videos and more.

    chatbots cocaine Recipes Researchers Share trick Wild
    Elan
    • Website

    Related Posts

    Daya launches tokenised US stocks for African investors

    July 23, 2026

    Kraken Parent Company Makes Moves to Offer xStocks From Global Markets

    July 22, 2026

    Winklevoss twins gave Trump’s super PAC $10 million 23 days after CFTC joined Gemini relief bid

    July 21, 2026
    Leave A Reply Cancel Reply

    Economy News
    Crypto

    Daya launches tokenised US stocks for African investors

    By ElanJuly 23, 20260

    Daya, a Nigerian startup building stablecoin-powered payment infrastructure for African businesses, has launched tokenised stocks…

    WayaWaya appoints ex-Chase Bank Kenya executive as board adviser

    July 23, 2026

    Reddit and Major Publishers Consider Blocking Google as AI Overviews Cut Search Referral Traffic

    July 23, 2026
    Top Trending
    Crypto

    Daya launches tokenised US stocks for African investors

    By ElanJuly 23, 20260

    Daya, a Nigerian startup building stablecoin-powered payment infrastructure for African businesses, has…

    Tech

    WayaWaya appoints ex-Chase Bank Kenya executive as board adviser

    By ElanJuly 23, 20260

    WayaWaya, the Kenyan company building AI-powered financial services for banks and merchants,…

    Tools

    Reddit and Major Publishers Consider Blocking Google as AI Overviews Cut Search Referral Traffic

    By ElanJuly 23, 20260

    Reddit and several major publishers are reevaluating the amount of access they…

    Your source for comprehensive insights on Africa’s private credit markets, InclusiFund synthesizes deal pipelines, repayment patterns, collateral trends, and sector-level signals to guide investors in underwriting and structuring credit in emerging African markets.

    We're social. Connect with us:

    our Categories
    • Work With Us
    • Advertise With Us
    • Contact Us
    • Disclaimer
    • Privacy Policy
    • Terms and Conditions

    Subscribe to Updates

    Get the latest creative news from FooBar about art, design and business.

    Facebook X (Twitter) Instagram Pinterest
    • Work With Us
    • Advertise With Us
    • Contact Us
    • Disclaimer
    • Privacy Policy
    • Terms and Conditions
    © 2025 Inclusifund. All Rights Reserved.

    Type above and press Enter to search. Press Esc to cancel.