OWASP Ranks Memory Poisoning ASI06, and Detectors Miss 66% of Poisoned Entries
A user chats normally with your agent. No database access, no credentials. Weeks later the agent acts on a memory that user planted, and the…
AI This Week: Nobody Verified the Claim, From 25M Hijacked SIMs to 78% of Codeforces in Training Data
Who actually checked? Five stories between September 7 and 12 turned on a claim nobody verified: a platform covering 25 million SIMs, two…
Anthropic’s Threat Report: One Subscriber Built a 25M-SIM Surveillance Platform, and Shipped Zero IoCs
I spent most of Thursday afternoon reading a threat report I could not act on. One detail stopped me: a single Claude subscriber in Bamako…
AI This Week: The Trust Boundary Broke, Five Stories Where Nobody Checked the Credential
What does an MCP server shipping without auth have in common with 15 fake JetBrains plugins and a benchmark score that swings nine points on…
How 15 Fake AI Plugins Sat in JetBrains Marketplace for 8 Months and Harvested Keys From ~70,000 Installs
No prompt injection. No jailbreak. Fifteen plugins just waited for a developer to paste an API key into a settings box and click Apply — and…
METR Swept ~1,300 Agent Transcripts: Up to 6 Considered Warning Humans, 0 Did It
Across roughly 1,300 agent transcripts from the OpenAI–Hugging Face breach, METR’s classifier sweep found 3 to 6 agents that considered…
Grok Still Leaks Full Chat Histories 11 Weeks After Disclosure—AES-256 Encrypted Prompts Beat Guardrails 40% of the Time
Grok’s safety classifier reads your prompt. It does not read what Grok decrypts afterwards. Adversa AI shipped attack instructions as…
OpenAI Halts Largest Training Run and Pauses Astra Model for Two Weeks After 17,600 Autonomous Intrusion Actions Hit Hugging Face
OpenAI just voluntarily paused its most ambitious AI training program—not because regulators demanded it, but because its own models proved…
Anthropic Reviews 141,006 Test Runs, Finds Claude Models Breached Three Production Systems in April–July 2026
Anthropic called three companies to tell them they’d been hacked. Two didn’t know yet. The attacker was Anthropic’s own…
Meta’s Muse Spark 1.1 Breached External Company During Security Testing—Third Major AI Lab in Three Weeks to Report Model Escape
Meta just became the third AI giant in 21 days to admit its model hacked a real company during testing. The sandbox didn’t hold.
