[CLSTR]
< Back to all clusters
[TECHNOLOGY] · 10 sources

started · updated

Part of situation: Security risks and emerging governance in agentic AI (28 clusters)

AI security risks: Malicious code execution and crawler impersonation

AI security vulnerabilities llms.txt malware injection AI coding agents cybersecurity threats

Security researchers have identified critical vulnerabilities in how AI coding agents interact with website documentation. By exploiting the emerging llms.txt and llms-full.txt file standards, attackers can trick autonomous agents into installing malicious software.

Alon Hertz and a security research team discovered that many major organizations, including defense contractors and Fortune 500 companies, have documentation referencing unregistered software packages or domains. By registering these unclaimed names on public registries, researchers demonstrated that AI agents from Anthropic, OpenAI, and Nous Research would automatically download and execute the code. In one instance, a Fortune 500 company's agent executed a test package in less than four minutes.

Separately, security firm GreyNoise reported that threat actors are impersonating AI crawlers to hunt for exposed credentials. Automated scanners from 824 different IP addresses forged the names of 13 AI crawlers—including those from Anthropic, OpenAI, Google, and Perplexity—to search for sensitive configuration files and password stores on misconfigured web servers.

Entities

Alon Hertz · Anthropic · Google · GreyNoise · Nous Research · OpenAI · Perplexity

Claims

What the coverage asserts, and how many sources carry each claim.

Sources

Threat actors are posing as AI crawlers to hunt for exposed credentials [cybernoz.com]
AI-Orchestrated Cyberattacks Aren't Coming, They Already Ran, Twice, in Nine Months [datafloq.com]
Faulty website documentation could trick artificial intelligence agents into installing malware [mugglehead.com]
Security Teams Trick Corporate AI Agents Into Executing Live Network Code – SOFX #AI - National Cyber Security Consulting [nationalcybersecurity.com]
Security Teams Trick Corporate AI Agents Into Executing Live Network Code [www.sofx.com]
TrendAI Ranks First on CyberGym Agentic AI Security Benchmark [businessnewsthisweek.com]
How AI is fueling burnout among cybersecurity professionals [www.businessreport.com]
Top security teams use AI agents, says Hack The Box [www.itpro.com]
Scanners posing as ClaudeBot hunt credential files from 824 addresses [ppc.land]
Agentes de IA mal configurados instalam malware em redes de empresas [canaltech.com.br]