Claude was asked to generate 50 "random" passwords. Results: only 30 unique. The most common one repeated 18 times (36% probability vs expected 2^-100). All start with uppercase letter + 7. Zero repeating characters within any password. LLMs literally cannot do randomness. Do not ask them for passwords.

Three papers on side-channel attacks against LLMs - all exploiting encrypted traffic metadata: - Infer conversation topics (medical vs coding) with 90%+ precision - Fingerprint queries with 100% accuracy via speculative decoding patterns - Identify sensitive topics like "money laundering" with 100% precision TLS is not enough.

NVIDIA is going to space. Space-1 Vera Rubin - AI data centers designed for orbit. Named after the astronomer whose work revealed dark matter. Cool stuff. nvidianews.nvidia.com/news/s…

JP Morgan's CISO wrote an open letter about agentic AI access management. AI agents are starting to act autonomously across systems, reading emails, executing code, calling APIs. Each one needs credentials, permissions, and guardrails. We haven't figured out how to manage human access properly (38% of ex-employees still have access). Now we're handing keys to autonomous agents. The access-trust gap isn't closing. It's about to get exponentially wider.

38% of former employees can still access their old company's data after leaving. SSO was supposed to solve this. But 34% of apps aren't even behind SSO. The offboarding checklist says "revoke access." Reality says there are 15 SaaS tools nobody in IT knows about, with passwords saved in browser autofill on a laptop that was never returned. We automated onboarding. Offboarding is still vibes.

The malware is invisible. Not obfuscated. Not hidden in a minified file. Actually invisible. GlassWorm used Unicode variation selectors, characters that are part of the Unicode spec but produce zero visual output. Your code editor would show blank lines. GitHub's diff view showed nothing suspicious. But the JavaScript interpreter would execute just fine. A VS Code extension with 35,800 downloads. Malicious code hiding in plain sight, or rather, hiding in plain nothing. Our previous security models assumed humans can review code. GlassWorm just proved that assumption wrong. And the C2 infrastructure? Solana blockchain. Memo fields on transactions carry the payload URLs. Immutable, anonymous, can't be taken down. Cost to rotate: $0.000005.

GLM 4.6, far from the most capable model out there, produced zero working smart contract exploits with basic prompting. Give it an agentic framework? 16 correct exploits. And here's the counterintuitive part: step-by-step instructions made things worse. Over-specified prompts overconstrained the agent. Tell it what's wrong, not how to fix it. The micromanager problem applies to AI too. Autonomy beats hand-holding. 32 working exploits from 23 real-world vulnerabilities. $135 total cost.

37% of employees knowingly break their company's AI policy. Not accidentally. Knowingly. Shadow IT was USB drives and unapproved Dropbox accounts. Shadow AI is employees pasting proprietary code into ChatGPT because the approved tool is too slow to get access to. 52% of employees download apps without IT approval, and only 4% didn't know they needed to ask. They know the rules. They just decided the rules aren't worth following. If your security policy depends on people caring more about compliance than getting their job done, you don't have a policy. You have a suggestion.