Show Notes
In this episode of Show & Tell, researcher Hal Triedman showed us how easy it can be to poison an AI research agent.
Hosts Craig Silverman and Alexios Mantzarlis talked with Hal, a computer science PhD student at Cornell Tech and former senior privacy engineer at the Wikimedia Foundation, about his preprint “Deep-Research Agents Can Be Poisoned via User-Generated Content.”
Hal walked us through his findings that demonstrated how you can trick AI research agents into citing fake products just by appending short text to Reddit pages. Hal also shared several free resources that can help you investigate Wikipedia.
What You’ll Learn: 📍 How AI research agents can be "poisoned": Adding a sentence or two to a Reddit page that Google or OpenAI deem relevant for a query can cause research agents to cite products, apps, and investments that don't exist. 📍What this means for AI search manipulation: At roughly two cents per run, a deceptive actor can test thousands of queries for a couple hundred dollars, with success rates of 20–40% on topics where UGC sites feature prominently on Google. 📍How to find stories on Wikipedia: Where to find pageviews for Wikipedia pages over time as well as how to follow discussions among Wikipedians about how to handle certain topics and controversies.
About Hal Triedman: Hal is a second-year PhD student in Computer Science at Cornell Tech, working with Vitaly Shmatikov’s lab on privacy, security, and AI. Prior to that, he worked as a Senior Privacy Engineer at the Wikimedia Foundation.
Links Referenced in Episode: 📍 Hal’s preprint — arxiv.org/abs/2605.24245 📍 Pageview data for Wikipedia – pageviews.wmcloud.org and https://analytics.wikimedia.org/publi... 📍 Wikimedia Enterprise API · The Signpost 📍 Follow Hal on Bluesky — harold.bsky.social
About Show & Tell Show & Tell from Indicator takes you inside digital investigations as leading journalists, OSINT investigators, threat analysts, and researchers walk through real cases they've cracked. Each episode features a recognized figure in the digital investigations community. They don't just tell us what they did — they show us.
Hit subscribe to never miss an episode
Follow on our other platforms for more content: 📍 Newsletter: indicator.media 📍 Bluesky: @indicator.media 📍 Instagram: @indicatordomedia 📍 LinkedIn: indicatormedia
More from Show & Tell from Indicator
In this episode of Show & Tell, we spoke to Shayan Sardarizadeh about his approach to debunking viral fakes for BBC Verify. "I want to check everything for AI," he told us. "I don't care where it comes from. If it comes from a national TV, if it comes from a politician posting it on their official TikTok, I don't trust anything." What You'll Learn: 📍Choosing where to look, and what to ignore when it comes to viral fakes 📍Why AI detectors are now an essential element of verification and why Shayan prefers the SynthID-based ones 📍How detectors can still fail, with a worked example from a viral image of Donald Trump 📍How reverse search is an essential complement to AI detectors and a mandatory starting point for any verification task Links mentioned in the episode 📍Shayan's thread on the viral Trump image: https://x.com/Shayan86/status/2100642920437108946 📍"AI generators are now required to offer detection tools. We tested them (and they need work)" (Indicator): https://indicator.media/p/ai-generators-are-now-required-to-offer-detection-tools-we-tested-them-and-they-need-work 📍You can follow Shayan on X at @Shayan86 and on LinkedIn: https://www.linkedin.com/in/shayan-sardarizadeh-0121555a/ About Shayan Sardarizadeh: Shayan is a senior journalist at BBC Verify covering disinformation, conspiracy theories and AI. About Show & Tell: Show & Tell from Indicator takes you inside digital investigations as leading journalists, OSINT investigators, threat analysts, and researchers walk through real cases they've cracked. Each episode features a recognized figure in the OSINT, digital investigations, or trust & safety community. They don't just tell us what they did — they show us. Hosted by Indicator co-founders Alexios Mantzarlis and Craig Silverman.
In this episode of Show & Tell, we speak to Renée DiResta about Grokipedia and the contestability of AI-mediated information sources. Grokipedia is worth studying not just as “a culture war artifact,” she said, but because even if Musk’s AI encyclopedia fades away, “something like it will exist at some point, and we may as well get in and figure out how to audit it now.”
In this episode of Show & Tell, hosts Craig Silverman and Alexios Mantzarlis sit down with Wall Street Journal reporters Caitlin Ostroff and Neil Mehta, who found that Polymarket paid creators to place fake bets on a faked version of its platform.
In this episode of Show & Tell, Craig and Alexios talk with Omer Benjakob, the disinformation and cyber reporter for Haaretz.
In this episode of Show & Tell, hosts Craig Silverman and Alexios Mantzarlis talk with Barbara Marcolini, a visual investigative journalist with Amnesty International.