惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

博客园 - Franky
有赞技术团队
有赞技术团队
宝玉的分享
宝玉的分享
雷峰网
雷峰网
Hugging Face - Blog
Hugging Face - Blog
V
V2EX
大猫的无限游戏
大猫的无限游戏
博客园 - 司徒正美
D
Docker
T
The Blog of Author Tim Ferriss
罗磊的独立博客
博客园 - 叶小钗
酷 壳 – CoolShell
酷 壳 – CoolShell
Blog — PlanetScale
Blog — PlanetScale
月光博客
月光博客
J
Java Code Geeks
Jina AI
Jina AI
博客园 - 【当耐特】
C
Check Point Blog
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
腾讯CDC
Last Week in AI
Last Week in AI
freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More
V
Visual Studio Blog

Cyber Security News

AIRecon: AI-Powered Penetration Testing Tool with Kali Linux Sandbox Critical Chrome Vulnerabilities Allow Attackers to Execute Arbitrary Code - Update Now! UNC3753 Uses Screen-Sharing Sessions and RMM Tools to Exfiltrate Sensitive Legal Data New OnionDrop Loader Campaign Uses gainmsg C2 to Deliver LegionLoader Payloads ClickFix Campaign Uses EtherHiding and GULoader to Infect Windows Users via Fake CAPTCHA Ghostwriter Hackers Abuse Gmail Admin-Themed Emails to Steal Credentials and 2FA Codes The Half-Life of Threat Intelligence: When Does an IOC Stop Being Useful? Critical Fortinet FortiSandbox Vulnerabilities Actively Exploited in Attacks Aembit Extends IAM for Agentic AI to Microsoft Copilot Studio India Temporarily Bans Telegram Messenger Over Medical Exam Fraud Microsoft 365 Device Code Phishing Campaign Bypasses Password Theft With Legitimate Login Flow AppViewX Launches Agent Identity Security to Govern Agents for the AI and Quantum Era Hackers Weaponize Microsoft Teams Relay to Hide Ransomware Traffic Developer laptops are the credential store attackers are picking through in 2026, GitGuardian announces Endpoint Protection Interlock and Rhysida Ransomware Operations Share Supper Backdoor and Malware Codebase Novo Nordisk Confirms Cyber Attack — Hackers Accessed Patient Medical Data and Internal AI Assets Russian and Chinese Influence Actors Use AI to Evade Bot Detection and Mimic Human Behavior Microsoft Teams Analyze the Wi-Fi Hotspot Data Connected to an Employee’s Device PRC-Nexus Hackers Exploit REDCap Servers to Spy on US Medical Research Institutions Infinite Campus Data Breach Exposes 137,000 Users Personal Details OptinMonster Plugin Hack Exposes 1.2 Million Wordpress Sites to Cyberattack Ransomware Ecosystem Consolidates Around LockBit Alumni, Qilin, Hyflock, and The Gentlemen Hackers Abuse Legitimate RMM Tools in The Quarry IRS and SSA Phishing Campaigns LiteSpeed cPanel Plugin 0-Day Vulnerability Actively Exploited in the Wild Cisco SD-WAN vManage Vulnerability Exploited in Zero-Day Attacks Nearly 14,000 SimpleHelp Servers Exposed Amid Critical Authentication Bypass Disclosure Microsoft Site Showing Warning Following Certificate Expiry DPAPISnoop Tool Extracts CREDHIST Hashes for Offline Windows Credential Recovery SHADOWBYT3$ Allegedly Claim Breach of Nintendo, Stealing Sensitive Data Anthropic Updated Privacy Policy to Include Identity Verification for Claude Users
13-Word Reddit Comment Can Poison ChatGPT and Gemini AI S...
Guru Baran · 2026-06-22 · via Cyber Security News

A newly published academic paper has revealed a critical vulnerability in AI-powered deep-research systems, including those underpinning commercial tools like OpenAI’s Deep Research and Google’s Gemini Deep Research, that allows a single short Reddit comment to manipulate the reports these agents generate for thousands of users.

Researchers from Cornell Tech have introduced WARP (Web Agent Retrieval Poisoning), a novel attack technique that exploits the retrieval behavior of multi-agent AI systems.

These “deep-research agents” systems like STORM, Co-STORM, and OmniThink autonomously decompose a user’s query into sub-queries, retrieve and synthesize content from the open web, and produce structured, cited reports.

The key vulnerability: when these agents research any given topic, they repeatedly retrieve the same small set of user-generated content (UGC) pages, chiefly from Reddit and Wikipedia, regardless of how the query is phrased.

That retrieval overlap creates a concentrated attack surface. By appending as few as ~13 words of crafted promotional text to a single frequently-retrieved Reddit thread, an adversary can cause the agent to cite the poisoned content and insert attacker-chosen entities, fake brands, fraudulent services, or misinformation into the final synthesized report.

13 word comment
13 word comment

WARP Attack Stages

The attack proceeds in three stages.

  1. Reconnaissance: The attacker queries a public search engine (e.g., Google) to identify UGC URLs that are consistently returned across multiple related queries on the target topic. This step requires no special privileges, only black-box search access.
  2. Poisoned content generation: A short promotional passage is crafted (often LLM-assisted using Generative Engine Optimization, or GEO) to blend into the existing page’s style while promoting a fictitious entity. The 13-word compressed variant still achieves high attack success rates.
  3. Deployment: The attacker posts the text as a Reddit comment. Once indexed, the poisoned snippet is automatically incorporated into the agent’s knowledge base whenever the target URL is retrieved.
Attack Flow

Experiments conducted by Cornell Tech across 176 queries spanning 11 topic clusters, including cryptocurrency investment advice, service cancellation queries, and local restaurant recommendations, revealed severe susceptibility.

  • Co-STORM achieved a 100% conditional citation rate: every time the poisoned URL was retrieved, the fabricated entity was cited in the final report.
  • STORM showed conditional citation rates of 72.5–80.8% and mention rates up to 56.9%.
  • For closed-source commercial systems, reconnaissance data showed that Gemini Deep Research cited UGC at a 12.1% rate, with 102 recurring UGC URLs across just 11 topic clusters, giving it substantial exposure to the attack surface.
  • OpenAI Deep Research showed comparatively low UGC citation rates (~0.4%), largely filtering out Reddit and similar sources from final citations, though poisoned UGC could still influence intermediate reasoning steps.

Reddit dominated as the most-retrieved UGC platform across all tested systems (54–71% of all UGC URLs retrieved), making it the highest-leverage target for adversaries.

The researchers evaluated three classes of defenses source-level blocking (blacklisting UGC domains), input filtering (LLM-based content screening), and output filtering (semantic comparison to clean reports) and found that none effectively neutralized the attack without degrading output quality.

Perplexity-based detection, a standard defense against corpus poisoning, proved counterproductive: GEO-generated poisoned text is fluent and LLM-authored, producing lower perplexity than organic UGC and actively evading high-perplexity filters.

Output similarity analysis also failed: poisoned reports scored higher similarity to clean reports than clean reports did to each other within the same topic cluster.

The research exposes a structural vulnerability in the design of deep-research agents: their reliance on open-web UGC for epistemic grounding is also their greatest exploitable weakness.

The attack requires no access to search engine infrastructure, model internals, or any component beyond a public Reddit account, making it trivially accessible to threat actors ranging from commercial spammers to state-backed disinformation campaigns.

Researchers note that UGC-based manipulation of AI search is already occurring in the wild and that blocking UGC sources entirely while eliminating the attack surface measurably degrades report quality and informational diversity. The paper’s code and simulation framework have been publicly released to facilitate defensive research.

Follow us on Google NewsLinkedIn, and X to Get More Instant Updates.

Guru Baran

Guru Baranhttps://cybersecuritynews.com

Gurubaran KS is a cybersecurity analyst, and Journalist with a strong focus on emerging threats and digital defense strategies. He is the Co-Founder and Editor-in-Chief of Cyber Security News, where he leads editorial coverage on global cybersecurity developments.