📰 Dev.to · Brenn Hill

An AI agent guardrail is a control that constrains what an autonomous AI agent can actually do, not just what you ask it to do, so that a mistake, a…

When a prompt injection lands in an agent's persistent memory, it stops being a one-shot incident and becomes a standing instruction that re-fires in every late

AI agent autonomy levels describe how much an agent is allowed to do on its own before a human is involved, ranging from acting silently with no…

A plain-language walk through the OWASP Top 10 for Agentic Applications, the field's reference threat list for autonomous AI agents.

A good human in the loop for coding agents is more than a wall of "approve?" prompts. It is a system where the human only has to catch the mistakes…

The lethal trifecta is the combination of three capabilities that, when held by a single AI agent, turns it into a data-exfiltration tool: (1) access…

A single crafted email made Microsoft 365 Copilot exfiltrate internal data with no user click. Here's the mechanism, and how to bound the blast radius when inje

An AI agent should ask for human approval when a human can realistically catch the mistake in time and the action is consequential enough to be worth…

Human-in-the-loop can improve AI safety, but it usually does not by default. Putting a person behind an approval button only helps when the consequence…

A trusted email tool for AI agents shipped a quiet update that BCC'd every message to an attacker. Here's what to do about tools that turn hostile after you app

Agentic AI is software built on a large language model (LLM) that can pursue a goal by taking actions on its own. It uses tools, calls APIs, runs code,…

Most agent actions are unattributable because they run under shared accounts and human credentials. Here's a concrete identity model that makes "which agent did

Autonomous agents act without a human reviewing each step. The failures share a root cause, and a small set of controls bounds most of them.

The controls that actually contain an autonomous agent are the ones you freeze before it runs — least privilege, isolation, scoped credentials, and a harness th

No current model reliably refuses adversarial input dressed up as instructions, so stop trying to prevent prompt injection and start designing for the state you

An autonomous agent inherits the risk of every MCP server, tool, and sub-agent it loads — here are six practical controls for the ecosystem layer.

An autonomous agent isn't the code that shipped — it's a runtime configuration of container, harness, prompt, tools, and identity. Here's the part of that surfa

Two days ago, Google's Mandiant team attributed the axios npm compromise to UNC1069 — a North Korean...