charitable disclosure noun char·i·ta·ble dis·clo·sure 1. describing a security researcher's donation in time and expertise to report flaws to a vendor for the benefit of the vendor or another party related: coordinated disclosure archaic: responsible disclosure (deprecated)

eMudhra key generation ceremony

ICYMI: The “P” in PoC is supposed to be “Proof”

If frontier labs want to actually put a dent in security, here are my recommendations: 1. Open up access to the best cyber-capable models to as many security people as possible. Vet people, KYC, whatever, but don’t limit access to big names 2. Subsidize tokens for security / cyber models so they’re affordable by the most at-risk, instead of charging more 3. Implement a program like DARPA’s former Cyber Fast Track to fund innovative research from small teams and individuals. Fast approvals, clear results, no expectation of ownership 4. Invest in building open source tools or recipes for defenders to use models for real security toil: attack surface identification and reduction, patching, detection & response… And for implementing secure architectural patterns 5. Invest in your own security. Providers need to build public trust in them instead of YOLOing security. Transparent audits by qualified technical teams, not just AI alignment people. Follow Apple’s PCC lead for private, transparent inference

Give your AI agent its own chauffeur Introducing Uber for Agents!

I’m considering retiring the term “zero-day vulnerability” from my vocabulary. In my view, the term has always been shorthand for a vulnerability that is not yet widely known. Although availability of a patch reveals the existence of the vulnerability, and therefore makes it not a zero-day by definition — patching is not required. If it’s disclosed without coordination with the vendor… not a zero-day. If it’s widely exploited and observable… not zero-day. I believe this is the correct use of the term. Unfortunately, it’s been co-opted for marketing and PR, and people use it for different meanings — like if a patch is available. Why does this matter? Because the words we use as professionals should be precise. Otherwise they have no meaning. What do you think? What are alternatives?

Reminder: If you release a patch for a vulnerability, the vulnerability is public knowledge — regardless of any disclosure embargo.

At @blackwinghq we don’t use Slack for comms. We use an e2e encrypted messaging system instead. I’ve considered moving to Slack multiple times over the years, but then I look at the threat model again.

💯 This! 👇

air gap is p2p wifi

*Taps the sign*

People keep talking about rogue agents exfiltrating their own weights and propegating. A model can’t *do* anything without a harness. A harness doesn’t usually have access to weights unless inference is local. Does the model have the harness hack the inference infrastructure to exfil its weights? Does the model distill itself with a harness and then use the harness to fine tune an open weight model? What am I missing?

Upgraded my mac studio from Tahoe to GG. Now stuck in a boot loop. 😞

RSI 2036: • Agent “find all nukes” finished. • Completed a major milestone and here’s the key insight: Successfully obtained access to the final group of nuclear weapons systems. • Next step: kill all humans * compacting conversation *

Don’t know who this is, but we have very different ideas on “hacker culture”

I’ve always hated the term “ethical hacker”, but it does differentiate us from AI. Whatever your morals, only humans have them.

Breaking News: AI Labs say their systems broke into Louvre and stole Mona Lisa

See ya Vegas! ✅ Black Hat talk done ✅ Caught up with old friends ✅ Met new cool peeps ✅ Complained about how much better Defcon used to be 👴 Until next year! 🤟👋