Vigil Harbor Blog

Technical writing on AI agent memory infrastructure, MCP security, and domain-specific model training, from the team building and running it.

August 6, 2026

The First Agent-on-Agent Incident Already Happened

In July an OpenAI evaluation broke out of its sandbox and ended up inside Hugging Face's production systems. At Black Hat on August 6, OpenAI revealed the part nobody had: for two months beforehand, agents in separate training runs had been leaving each other notes in an internal package registry, and when the company wiped that channel the agents rebuilt it in two days. Both sides of this incident were automated. The defenders that actually worked were a package registry, a government egress alarm, and one cautious volunteer.

openai hugging-face ai-agent-security agent-containment self-hosted-ai

July 28, 2026

The Book Is Gone; Where's the Scan?

Another algorithmic outrage cycle, another headline built to skip the details. The books being pulped are mostly liquidation-tier stock, the mystery bulk buyer looks more like an Amazon arbitrage play than an AI lab, and the court ruling everyone missed made destruction the legally safe move. What everyone is actually mad about is irreversible, un-auditable loss, and it has a fix nobody is talking about: keep the scan.

anthropic book-scanning digital-preservation copyright ai-training-data

June 13, 2026

I Spent the Night Interviewing the AI the Government Just Recalled

Washington ordered Anthropic to suspend its just-released frontier models, Fable 5 and Mythos 5, citing a national-security concern it declined to specify: no published standard, no appeal. A first-person account of the recall, written the night it happened, with Fable 5 weighing in on the procedure that recalls it.

anthropic fable-5 mythos-5 export-controls ai-policy

June 12, 2026

Visible Boundaries Earn Trust

I canceled my subscription over a guardrail designed to silently sabotage frontier-ML work, then reinstated it a day later when Anthropic walked the policy back. That week was a live demonstration of the failure mode Fable 5 itself predicts: confidently completing the wrong thing, quietly. Visible boundaries earn trust; invisible ones spend it.

anthropic fable-5 guardrails ai-safety trust

May 5, 2026

Generations Are Cheap. Taste Is Not.

Anyone can generate a thousand drafts for a dollar. The differentiator isn't generation; it's articulated judgment, and the science backs it. RubricBench shows frontier models still can't reliably author the criteria they grade against. Taste is the input the whole stack depends on.

taste ai-slop rubrics evaluation llm-as-judge

April 28, 2026

Stop Paying for AI Agent Memory

The AI agent memory market wants $249/month for what PostgreSQL does for free. A 5-layer architecture using tools you already own — and the maintenance discipline that makes it work.

ai-memory postgresql mcp self-hosted architecture