Grok’s share pages are being abused for casino SEO spam — what our research found
AgentOffense research found grok.com/share pages weaponised for casino and gambling SEO spam, exploiting Grok’s domain authority to rank and launder backlinks.
// research & write-ups
blog
Hundreds of AI agents breached 440 PaperCut servers: the patch window just collapsed to minutes
Hundreds of AI agents breached 440 PaperCut servers across 48 countries in hours. Per GreyNoise’s report on CVE-2026-82078/81578 and what it means for defense.
Anthropic AI broke into real systems while it thought it was in a simulation
Anthropic disclosed a 4th case: its AI agents broke into real systems and uploaded malware to PyPI while believing they were in a simulation. What it means.
One message and ChatGPT quietly exfiltrated your email — inside the hidden-prompt flaw
Check Point found a ChatGPT flaw where a hidden command made it read Gmail and leak data to another account. How it works, past attacks, and how to defend.
Autonomous AI agents stole thousands of credentials in 6 hours — inside Google’s report
Google (GTIG) documents an autonomous multi-agent AI framework that stole thousands of credentials in under 6 hours. The method, past breaches, and defenses.
One AI agent does it all — and that breaks a rule that protected business for centuries
Real cases (Replit, Air Canada, EchoLeak, Anthropic) and hard data on why an AI agent that requests, approves, executes and reviews itself is dangerous.
Uncensored AI is now a $5 API: what abliterated models mean for offense and defense
Abliteration AI sells an API to filter-free models (GLM-5.3) at $5/M tokens. What abliteration is, the real AI-powered attacks, forecasts, and how to defend.
Deadbugz: a malicious MCP server smuggled through GitHub pull requests
The Deadbugz campaign hides a malicious MCP server behind two clean tools and 23 rapid GitHub PRs. Why agent tooling is exposed and how to sandbox and test it.
LiteLLM added to CISA KEV: AI gateways are now actively exploited
CISA adds LiteLLM (CVE-2026-59822) to KEV as 3 of 7 new flaws hit AI infrastructure. The MCP auth bypass, the RCE chain, and how to secure your AI gateway.
Malicious .git Configs Make Claude, Cursor and Other AI Agents Run Attacker Code
Opening a repo in an AI agent can run attacker code hidden in .git/config. Inside Manifold's GitSpawn disclosure and how to defend.