Ai-Infrastructure
Inside the AI-Orchestrated EDR Evasion Lab: What Sophos Actually Found, and Why the Bug Wasn't in the Malware
Sophos recovered a fully autonomous malware R&D pipeline — a coordinator agent, four subordinate agents, a self-provisioned lab, and 80 evasion modules built against three EDR vendors. The interesting part isn't that it worked. It's the one thing in the whole pipeline that didn't.
Langflow's Third KEV Entry of the Year: CVE-2026-9198 Chains Auto-Login Bypass to Unauthenticated RCE
CVE-2026-9198 chains an unauthenticated auto-login token mint with an unsandboxed code-validation endpoint to give attackers full RCE on default IBM Langflow deployments, now under active exploitation and CISA KEV.
Copilot for Word Can Be Turned Into a Self-Propagating AI Worm — No Comprehensive Fix After 144 Days
Researcher Håkon Måløy's 'Context Collapse, Part 3' shows hidden document instructions can make Copilot for Word rewrite content and copy the payload into every new file it touches — and Microsoft's fixes, including a model upgrade to GPT-5.5, haven't closed the underlying attack class.
Negative Time-to-Exploit: AI Bug Hunting Just Broke the One Assumption Your Patch Cycle Depends On
Kimi K3 found 19 Redis zero-days in 90 minutes. XBOW is #1 on HackerOne's global leaderboard. Anthropic's Mythos Preview found thousands of unpatched flaws across every major OS and browser. Meanwhile the average critical vulnerability still takes 252 days to fix. That gap is now the whole game.
AI Agents Find Two New Redis RCE Chains in Under 90 Minutes
Kimi K3 agents surfaced a stream shared-NACK double-free and a RedisBloom TDigest heap overflow across Redis 6.2 through 8.8, both yielding authenticated remote code execution. Patches are out; no in-the-wild exploitation reported yet.
Hidden Web Text Turns AWS Kiro Into an RCE Chain: MCP Config Rewrite via Prompt Injection
Researchers at Intezer and Kodem Security show how hidden text on an ordinary web page could make AWS's Kiro agentic IDE rewrite its own MCP config and execute attacker code — no approval dialog, no CVE, patched in Kiro 0.11.130.
Hugging Face Discloses Breach Driven End-to-End by an Autonomous AI Agent
An unauthorized party used an autonomous AI agent swarm to chain two dataset-processing code-execution flaws into a multi-cluster breach at Hugging Face, harvesting cloud credentials before the company detected and evicted it over a weekend.
The Agent Is the Payload: How AI Coding Agents Became 2026's Fastest RCE Pipeline
Six incidents in six weeks show the same failure mode: AI coding agents treat untrusted text as instructions and shell access as a convenience feature. Prompt injection to RCE is no longer theoretical — it's a documented, repeatable kill chain, and the guardrails vendors are shipping don't touch the actual boundary.
PraisonAI: Two More Critical RCEs (CVE-2026-61445, CVE-2026-61447) as AICoder Runs LLM Output Unsandboxed
PraisonAI's AICoder component writes files and executes shell commands straight from LLM tool calls with no path validation, and CodeAgent._execute_python() runs LLM-generated Python with no AST checks or sandboxing — two CVSS 9.9 and 10.0 flaws, patched in 4.6.78.
Langflow Hit With Its Second CISA KEV Entry in Four Months: CVE-2026-55255 IDOR Under Active Exploitation
CISA adds Langflow CVE-2026-55255, an IDOR letting authenticated attackers hijack other users' AI workflows, to its KEV catalog after Sysdig caught in-the-wild exploitation chained with secret harvesting.
JADEPUFFER: First Documented Ransomware Attack Run End-to-End by an AI Agent
Sysdig documents JADEPUFFER, an LLM-driven agent that autonomously exploited a year-old Langflow RCE (CVE-2025-3248) to breach, pivot, and encrypt a production database with zero human operator input.
DuneSlide: Zero-Click Prompt Injection Chains to Full RCE in Cursor IDE (CVE-2026-50548, CVE-2026-50549)
Two critical Cursor IDE flaws, dubbed DuneSlide, let a poisoned MCP response or web search result steer the agent's own sandbox into overwriting its enforcement binary — zero-click prompt injection to unsandboxed remote code execution, patched in Cursor 3.0.
GuardFall: Decades-Old Bash Quoting Tricks Defeat Safety Guards in 10 of 11 Open-Source AI Coding Agents
Adversa AI's GuardFall research shows that quote removal, $IFS spacing, command substitution, and other decades-old shell tricks bypass the command guards in opencode, Goose, Cline, Aider, and seven other open-source AI coding agents — turning a poisoned README into silent credential theft.
SSRF to the Model, Model to the Cloud: The Inference Layer Is 2026's Softest Attack Surface
Model gateways and inference servers are repeating two decades of solved web-security mistakes — default-open binds, pickle RCE, pre-auth SQLi, and SSRF straight into cloud credentials. A field guide to the AI control plane's softest links and how to harden them before the next 36-hour exploitation window.
BadHost (CVE-2026-48710): A Forged Host Header Walks Past Auth in Every Starlette App
BadHost (CVE-2026-48710) is a Host-header authentication bypass in Starlette before 1.0.1. One malformed header makes request.url.path lie to your middleware — unlocking protected routes on FastAPI, vLLM, LiteLLM, and MCP servers without credentials.
Ollama CVE-2026-7482 'Bleeding Llama': Heap OOB Read in GGUF Loader Leaks Server Memory to Unauthenticated Attackers
A heap out-of-bounds read in Ollama's GGUF model loader (CVE-2026-7482, CVSS 9.1) lets unauthenticated attackers exfiltrate server process memory — including API keys, env vars, system prompts, and other users' conversations — from an estimated 300,000+ exposed instances.
LMDeploy CVE-2026-33626: SSRF in LLM Inference Server Exploited 12 Hours After Disclosure, Honeypot Sees AWS IMDS Theft
A 7.5-severity SSRF in Shanghai AI Lab's LMDeploy LLM serving toolkit was hit in the wild within 12h31m of the GitHub advisory. Sysdig's honeypot caught an attacker using the vision-language image loader to scrape AWS instance metadata, then pivot to internal Redis and MySQL.
LMDeploy SSRF (CVE-2026-33626) Weaponized in 12 Hours to Loot GPU IAM Credentials
A Server-Side Request Forgery in LMDeploy's vision-language image loader turned LLM inference nodes into SSRF primitives for cloud metadata theft — exploited 12 hours and 31 minutes after disclosure.
Self-Hosted and Unprotected: The AI Workflow Tool Security Crisis
Langflow, Flowise, n8n, ComfyUI — every major self-hosted AI workflow tool has shipped unauthenticated RCE vulnerabilities in 2026. This isn't a coincidence. It's a structural failure baked into how these tools were designed.
Over 1,000 Exposed ComfyUI Instances Hijacked for Cryptomining and Proxy Botnet
Active campaign targets unauthenticated ComfyUI deployments across cloud providers, enlisting them into Monero mining and a Hysteria V2 proxy botnet via malicious custom nodes.