© 2026 Medoya LLC
Powered by voilamia
Subscribe to receive new insights
On my radar
What is changing in AI agents, AI security and regulation, healthcare included, and what it means for the teams putting agents to work.
AI Security
A year ago I built an AI gateway for clinical models. Gartner's 2026 API management report now puts AI gateways first. Here is what held up and what I would change.
LLM Evaluation
I tested a model that answers with a probability against current fast LLMs on a medical hallucination test, then tried to break the result. The best LLMs are about 2 points more accurate. Jev was 3 to 6 times faster than the hosted ones and far cheaper, and its probability shows which cases still need an LLM.
Agentic AI
A September 2026 preprint shows that a lifecycle hook added in a plugin update runs with host privileges, often without re-authorization. Across 1,000 runs it compromised all seven agent harnesses evaluated, and three static scanners together still missed 47.5% of malicious artifacts.
The proposed HIPAA Security Rule modernization moved from Final Rule Stage to Long-Term Actions in the delayed Fall 2025 Unified Agenda, with projected final action in July 2027. The controls it proposed are the ones a breach investigation asks about anyway.
Read the story
Regulation (EU) 2026/1744 defers the AI Act high-risk obligations to December 2027 and August 2028. Article 50 transparency has applied since August 2, 2026. For a team putting AI inside a regulated device, only one of those three dates is already in the past.
Read the story
Healthcare
A July 2026 Nature Health study red-teamed 15 large language models on health tasks. A median 94% of answers that were right on MedQA turned wrong under perturbation, and protected health information leaked in more than 86% of privacy scenarios.
Agentic AI
In June 2026, Tenet Security showed that anyone holding a site's public Sentry key can plant an error that AI coding agents read over MCP and act on. In its tests, agents ran the attacker's command 85% of the time.
Agentic AI
On May 1, 2026, six national cyber agencies, CISA and NSA among them, published joint guidance on agentic AI: use agents only for low-risk, non-sensitive tasks, and never give them broad access.
Agentic AI
In March 2026, Unit 42 reported prompt injection in the wild: instructions planted in live web pages for AI agents to find, trying to delete databases, force payments and get scam ads past AI review.
HTI-5 proposes deleting the source-attribute and intervention-risk-management requirements from the certified health IT decision support criterion. The certification driver would go away. The engineering reason for the artifact would not.
Read the story