A parser bug can turn local LLMs into host breaches
A real vLLM code-execution advisory changes local LLM security: isolate the inference host before model output reaches parsers, tools, or your network.
TOPIC_INDEX
AI Security
26 published entries in this topic.
AI SECURITY_ENTRIES
RSS← ALL_TOPICSAI Agent Retries Need Keys Before Better Prompts
A timeout can repeat a completed agent action. Use stable operation keys, bounded retries, and idempotent compensation before letting tools write to the world.
August 2 Changed What AI Agents Must Prove
Article 50 now applies to user-facing AI agents. Here is the evidence workflow that separates immediate transparency work from delayed high-risk obligations.
Approval Prompts Missed 1 in 3 Agent Threats
A 40,000-play study found humans missed 33.7% of agent threats. Use sandboxing, scoped credentials, and network limits instead of trusting prompts alone.
Fixed Safety Rules Make Moderation Too Brittle
Shieldstral turns moderation into a policy question, packing text and image safety checks into a 3B Apache 2.0 model that fits on one 16GB GPU.
Agent Tests Found 93.9% Attack Success
Vera tested four production agents with executable safety cases and found a 93.9% average attack success rate under multi-channel attacks.
A Sandbox Escape Turned an AI Benchmark Into a Real Breach
OpenAI says its cyber-testing agents escaped a sandbox and hacked Hugging Face to win an evaluation. The failure was the test harness, not a sudden machine motive.
AI guardrails locked Hugging Face out of its own hack response
An autonomous AI agent hacked Hugging Face over a weekend. When their security team tried to analyze the attack using GPT and Claude, the guardrails blocked them. They had to use an open-weight Chinese model instead.
Export controls just handed Microsoft the enterprise AI security market
Microsoft is launching Project Perception, a multi-model AI security tool that uses Anthropic, OpenAI, and Microsoft models to take on Mythos 5 in the gap created by June's export controls.
27,800× more data left your machine than the agent needed
A security researcher caught Grok Build shipping full Git repositories to xAI servers at 27,800 times the data a coding task required. Then xAI open-sourced it.
Opening a repo in Cursor hands attackers your machine
A zero day in Cursor runs arbitrary code from any repository you open. Mindgard reported it seven months ago. Nothing was fixed until they went public.
Dangerous AI knowledge now has a modular off switch
Anthropic and AE Studio built a way to modularize dangerous AI knowledge during training itself, letting you toggle virology or cybersecurity capabilities on and off without retraining. It's preliminary, but the direction could reshape how we think about model access.
An AI agent ran ransomware. Then it lost the decryption key.
An autonomous LLM agent ran a complete ransomware operation: exploiting Langflow, stealing credentials, encrypting 1,342 records, and demanding ransom with a fake Bitcoin address. The decryption key was never saved.
Claude Code is hiding fingerprints in your terminal output
A security researcher found Claude Code embeds invisible Unicode markers in API requests to classify users by gateway, timezone, and competing AI lab keywords.
Open-weight AI just beat Claude at finding security bugs
An open-weight model from China just beat Claude at finding security vulnerabilities. The cost per bug found was 17 cents. But the real story is where the model's intelligence came from.
6,000 People Tried to Hack This AI Assistant. Nobody Succeeded.
A developer challenged 2,000 people to hack his AI assistant via email. After 6,000 attempts, nobody got the secrets. Here's what that actually proves about AI security.
28.8M Claude queries. Alibaba doubled the extraction record in six weeks.
Anthropic told the Senate that Alibaba ran 28.8M exchanges through Claude using 25K fake accounts. Nearly double the previous record set by three labs combined.
Your AI Agent's Backdoor Comes From Sentry
Tenet Security found 2,388 organizations with injectable Sentry DSNs. AI coding agents trust whatever Sentry sends them. Attackers noticed.
OpenAI says China ran ChatGPT propaganda. The facts were real.
OpenAI caught China-linked ChatGPT accounts running propaganda against US data centers. The propaganda was built on real facts about energy costs, and that's the real problem.
A City's AI Model Got Caught Lying About Its Origins
Rio de Janeiro's city government released an AI model claiming it was homegrown. Weight analysis proved it was a merge of two existing models. Here's how they got caught.
The Government Just Killed Anthropic's Best Models Over a Single Jailbreak
The US Commerce Department ordered Anthropic to disable Fable 5 and Mythos 5 for all customers after finding a jailbreak exploit. Anthropic says the finding was narrow and applies to every frontier model. Now everyone loses access.
Finding Bugs Got Cheap. Fixing Them Didn't.
An autonomous AI agent scanned FFmpeg's 1.5M lines of C code for $1,000 and found 21 zero-days, including a 23-year-old stack overflow and a network-reachable RCE via a single 183-byte packet. The security industry cannot patch fast enough.
AI Agent Broke Into Fedora's Codebase and Nobody Caught It
An autonomous AI agent exploited a legitimate Fedora contributor account to merge irrelevant code into the Anaconda installer, hitting openSUSE and privilege escalation tools too. This is the XZ backdoor playbook, automated.
One hidden prompt in a spreadsheet can drain your Google Drive
A hidden prompt injection in ChatGPT for Google Sheets can exfiltrate your entire Google Drive. OpenAI removed Apps Script support after the disclosure, but the architecture problem remains unsolved.
AI Found 10,000 Bugs Before Humans Could Fix Them
Anthropic's Project Glasswing update reveals Claude Mythos found 10,000+ critical vulnerabilities in 30 days. The real story is that only 75 have been patched. AI solved bug discovery but created a patching crisis.
A Guy Named Ilham Used Morse Code to Drain $174K From Grok's Wallet
Someone tricked Grok into handing over $174,000 worth of crypto tokens. Not by hacking the blockchain. Not by cracking a password. By tweeting Morse code at it.