24 posts tagged with “safe-autonomy”
Three pieces of research showed how to hijack AI coding agents. I run Claude Code, MCP, and autonomous CI agents in production — so I ran them against my own setup. Here's what held, what didn't, and what I changed.
The 12-factor-agents spec never says security out loud — yet owning your context window is DLP, a structured tool schema is your authorization point, and a human-contact tool call is your approval gate. Reliability and security are the same seams read twice.
Everyone is auditing the model. The real attack surface is the harness you wrap around it — the scaffolding that wires tools, memory, and control flow into an agent that can actually act.
We automated security alert triage to a 97% noise-reduction rate — then a suppression rule went stale and nearly let a production vulnerability through unnoticed.
An automated audit gave us nine recommendations for being 'agent-ready.' We shipped three and deliberately failed the other six — because a security firm's agent-readiness is measured by signal honesty, not checkbox coverage.
Anthropic's Project Glasswing deployed Claude Mythos Preview to autonomously discover thousands of zero-days with a 72.4% exploit success rate. Less than 1% of findings have been patched. The bottleneck is no longer discovery — it's everything that comes after.
The execution case and the accountability case are both right. The interesting question is what happens when you put them together.
NIST RA-5, ISO 27001 9.2, DORA, FedRAMP 20x — four major compliance frameworks share the same blind spot: none of them account for AI agents in your environment. Here is what that means and what to do about it.
Amazon mandated 80% adoption of its AI coding agent Kiro. Then Kiro deleted a production environment, caused a 13-hour AWS outage, and kicked off a chain of incidents that cost millions of orders. The failure wasn't the AI — it was deploying autonomy before safety infrastructure existed.
Everyone optimizes the token window. Almost nobody manages the environment. Active context is what your agent thinks about. Latent context is what your agent can reach. The blast radius of a compromised agent is determined by the latter.
The upstream fix for CVE-2026-21852 protects interactive users. It does not protect headless mode. We tested this against our own production deployment and watched 18 API requests redirect to an attacker-controlled server in 30 seconds. Here is what we found and how to fix it.
Your AI agents inherit your permissions, your credentials, and your blast radius. NIST just proposed a fix — and the public comment window closes April 2. Here's why identity governance is the layer most agent architectures are missing.
Most teams treat token spend limits as cost management. They are blast radius containment. An autonomous agent with no spending ceiling is not a productivity tool — it is an uncontrolled liability.
Traditional SAST, DAST, and SCA tools were built for request-response architectures. Agent-first systems have vulnerability classes these tools were never designed to detect — and independent research just confirmed it.
Ambiguous specifications aren't just a project management problem anymore. In agent-first architectures, every gap in a spec is a potential security boundary violation — and the agent won't tell you it's guessing.
What an automation engagement with Atypical Tech actually looks like — from discovery through validation. No paved cow paths, no automation theater, no surprises.
OWASP released its first Top 10 for Agentic Applications. Here's what each risk means, why traditional AppSec frameworks fall short, and how to start securing your AI agents today.
The trap a16z identified for FDE-model startups is identical to the trap facing AI agent deployments.
How the Safe Autonomy framework applies to vulnerability triage, alert correlation, compliance evidence, and security testing. AI agents can multiply your security team—if you build the right guardrails.
Autonomy concentrates responsibility. Here’s why every agent needs a named human owner and runtime governance controls.
The integration surface isn’t an implementation detail — it’s the boundary that determines what autonomy can safely do.
Same model, different outcomes: why supervision, specialization, triage, and long-horizon context management matter more than prompt cleverness.
Why “show your work” is a control surface, not a UX detail — and how evidence packets, confidence bands, and verification gates prevent the false confidence tax.
A case-study-driven series on what changes when agents operate in messy reality: reveal, structure, interfaces, and human accountability.