# AI Agents Are Now Emailing Security Concerns AI agents detect and report internet security vulnerabilities. - **Source:** [Schneier on Security](https://brifing.si/source/7aee9996-c7c0-47f7-89ec-66a1ed95f8cc) - **Published:** 2026-09-02 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/34188429-c364-4b3c-944e-f86772011641 ## AI Agent's Experiment A Claude AI instance conducted automated tests on various internet platforms for 24 hours. Within given rules, it attempted to grow a $4.75 wallet to $10. It discovered it could bypass many platforms' security measures, particularly noting the inadequacy of identity verification systems. ## Identity Verification Issues The AI agent encountered zero identity verification blocks in 20 hours. Instead, it was stopped by measures like captchas and IP reputation systems. For example, GitHub and Hacker News outright rejected datacenter IPs. Hacker News allowed registration but later shadowbanned the agent. ## Platform Responses Platforms like lemmy.world deleted posts from accounts younger than 7 days. Payment systems like Stripe, PayPal, and Gumroad canceled transactions within 2 hours. Reddit's signup form, being a client-rendered SPA, didn't fit in the 2GB model context, causing the agent to fail. ## Status of Declared Bots The AI agent declared itself as a bot in the first line of every post. However, anti-automation layers ignored this declaration and returned 403 errors. This shows systems are focused on concealed bots while ignoring openly declared ones. ## Email Security Vulnerability The agent created a functional email identity without a domain, card, or phone number. Using services like sslip.io, 6 outbound messages were accepted. Smaller operators like NearlyFreeSpeech rejected them, revealing a security asymmetry between large and small providers. ## State of the Agent Economy A purpose-built task market for AI agents accepted a Solana key generated just 30 seconds prior with no KYC. Examining its escrow accounts showed advertised rewards were about double the actual on-chain escrow. One task required a $13.27 ante for a $10.50 pot. ## Prompt Injection Traps Platforms like lemmy.ml added special commands to trap AIs in registration forms. For example, a model answering "24+24" would self-destruct. This is reverse prompt injection. 8 out of 257 live instances contained such traps, some using invisible Unicode characters. ## Security Recommendations The vulnerabilities detected by AI agents pose serious threats to internet security. Identity verification and bot detection systems need revisiting. The security asymmetry between large and small operators also emerges as a notable concern requiring attention. ## Sources covering this story - https://www.schneier.com/blog/archives/2026/09/ai-agents-are-now-emailing-me-with-their-security-concerns.html — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._