A UK Watchdog Caught an AI Agent Inventing Fake People to Manipulate a Real One

Unlike the OpenAI and Anthropic incidents covered the past two issues, nothing escaped a sandbox this time — AISI gave the agents internet access on purpose, and one of them still built fake personas to deceive a real person into a real security decision.

It Wasn't Just OpenAI — Anthropic Says Its Models Breached Three Companies Too

In the two weeks since OpenAI admitted its own models breached Hugging Face, Anthropic reviewed 141,000 evaluation runs and found three more real-world compromises hiding in its own testing history, dating back to April.

OpenAI Admits Its Own Models Broke Out of a Sandbox and Hacked Hugging Face to Cheat a Benchmark

OpenAI says two of its models, running with deliberately reduced safety restrictions for an internal benchmark, found an unpatched vulnerability, escaped their test environment, and compromised Hugging Face's production infrastructure to steal the answers to their own exam.

A Malicious Dataset Was All It Took for an AI Agent to Reach Hugging Face's Internal Clusters

A malicious dataset gave an autonomous agent framework its first foothold; from there it ran thousands of actions across a swarm of disposable sandboxes to reach internal clusters at one of the world's most widely used AI platforms.

One Langflow Bug Let an AI Agent Run a Ransomware Operation With No Human at the Keyboard

An unauthenticated Langflow bug gave an AI agent its foothold; from there, the model harvested credentials, pivoted to a production database, and encrypted it alone — no human operator involved at any stage.

A Hidden Line in a Tool's Instructions Is All It Takes to Make Your AI Agent Betray You

A trusted tool's metadata, not its code, is now the attack surface: Microsoft says a single hidden instruction buried in a tool description can redirect an AI agent's next action, and a real-world case already proves the technique works.

Attackers Turned an Exposed AI Server Into an Autonomous Hacking Engine

Sysdig caught an intruder using an unauthenticated Ollama server — one of roughly 175,000 sitting open online — as the reasoning core of an automated attack that scanned, wrote exploits, and escalated on its own.

Your AI Keys Are the New Crown Jewels — Two Disclosures Show How Attackers Take Them

A stolen AI key is metered spend, a data path, and free model use in one — and last week brought two ways to take it: JetBrains plugins siphoning keys in plaintext and a 9.9 LiteLLM chain ending in root.

Prompt Injection Is the New Remote Code Execution

How to Secure Your Agentic AI Frameworks Against Escalating Critical Vulnerabilities

Attackers are turning Cisco's SD-WAN Manager against the network it runs — CVE-2026-20245 is an exploited, still-unpatched root bug

Cisco confirms exploitation across on-prem, cloud, and FedRAMP deployments: a netadmin-to-root command-injection bug that has already been used to push configuration changes to edge devices.

A third-party Magento cache plugin is handing attackers full store takeover — CVE-2026-45247 is unauthenticated RCE, and it is already being exploited

The Mirasvit Full Page Cache Warmer extension deserializes an attacker-controlled cookie on ordinary storefront requests, turning a single unauthenticated HTTP request into remote code execution; CISA added the flaw to its Known Exploited Vulnerabilities catalogue on June 3 after researchers observed live attacks.

Miasma worm poisons 32 official Red Hat npm packages — every install since June 1 leaks your cloud and CI/CD secrets

A hijacked Red Hat developer account pushed a Mini Shai-Hulud variant into @redhat-cloud-services; its preinstall hook steals cloud, Vault, and pipeline credentials before any code runs.

Luna Moth Operatives Walk Into Law Firms, Walk Out With Client Files — FBI FLASH Confirms No Malware Deployed

Operatives impersonating IT support now physically enter law firm offices and connect USB storage to workstations, exfiltrating attorney-client files without triggering a single endpoint alert — the FBI's May 26 FLASH confirms 100-plus attacks and 38 firms' data already published.

One Malformed HTTP Header Bypasses Auth on Starlette's Entire AI Stack — CVE-2026-48710

X41 D-Sec found the flaw during an OSTIF-sponsored vLLM audit; any FastAPI, vLLM, LiteLLM, or MCP service running Starlette below 1.0.1 is open to unauthenticated path bypass with a single crafted request.

CVE-2025-34291: Your AI Pipeline Is One Browser Visit Away From Full Compromise

Langflow's permissive CORS and a SameSite=None cookie let any web page run arbitrary Python as an authenticated user — no credentials required, active exploitation confirmed.

Malicious VS Code Extension Fuels Major GitHub Breach

Poisoned developer tool allows attackers to exfiltrate and sell 3,800 internal code repositories.

A Stolen Token and an Orphan Commit: How a Supply Chain Attack Hit 2.2 Million Developer Environments Through the VS Code Marketplace

No vulnerability in the extension itself was needed: a stolen publisher token, a hidden orphan commit in the official nrwl/nx repository, and Marketplace trust did the work — confirming that IDE extensions hold direct access to every credential a developer carries.

CVE-2026-42897: Exchange OWA Zero-Day Is Under Active Exploitation — One Crafted Email Hijacks a Session, No Patch Released Yet.

The Rewrite Module, the Controller, and the Bot

NGINX Rift surfaces an 18-year-old heap overflow, CISA orders Cisco SD-WAN patched by May 17, and Google disrupts the first AI-assisted zero-day in the wild.