NEAR Protocol (NEAR)

near(.)com quietly became a full terminal. spot across 35+ chains, a confidential mode if you don't want to be watched, and now perps routed through hyperliquid. same tab, one login.

$NEAR's front end used to be a swap box.

the swap box grew into a desk. https://t.co/NBtnJdqY5l

Peekaboo, old friend

Bought back some $NEAR this morning, first time touching it after TP'ing at $3 a few weeks ago

few more bids lower https://t.co/kOLswXmYm5

Last week I presented at the Real World AI Security Conference at Stanford hosted by @danboneh.

There were many papers presented on ways to attack LLMs and agentic systems, mostly leveraging prompt injection. Most research is trying to address prompt injection in the LLM by training itself, but I think this will be increasingly challenging and continue leaving probabilistic gaps.

We can’t blame LLMs for these vulnerabilities – any system can be attacked, including a human intelligence (e.g. phishing and ads).

We need to build AI systems with this threat vector as an assumption, rather than an exception. Every input is potentially an attack surface.

At @near_ai, we’re building @IronClawAI with exactly this systems thinking – assuming external inputs are compromised and designing a harness to handle that.

Broadly I think of the harness as an operating system, and we need to build more effective and secure AI OS that can be trusted with all of our context, sensitive data and decisions.