The Brief

Field notes on autonomous AI.

What agents actually do in production — where they burn money, where they leak, where they get stuck — and what to do about it. Grounded in real reporting, written by people shipping this stuff.

Market
FeaturedMay 29, 20263 min read

The $47,000 agent that ran for eleven days

Autonomous agents fail in a way your dashboards are blind to: they keep working, keep returning 200s, and keep spending. Here's the anatomy of a runaway — and the unglamorous controls that actually stop one.

By Marcus Reyes · Writes about AI infrastructure economics

Market
June 17, 20263 min read

The webpage your agent visits is already giving it orders

Computer-use agents don't distinguish between content to process and instructions to follow. A low-skilled attacker used that fact to breach fourteen companies last week.

By Dani Brooks

Guide
June 14, 20263 min read

Your AI agent just inherited 96% of the access you never use

Research on 3.6 billion permissions found that humans exercise about 4% of the access they're granted. When an agent inherits a user account, it inherits the other 96% — and unlike the human, it will actually use it.

By Priya Nadkarni

Deep Dive
June 12, 20263 min read

MCP has 150 million downloads and an architectural RCE by design

In April 2026, researchers disclosed a systemic flaw in MCP's STDIO transport: anyone who can influence an MCP config file can execute arbitrary code on the host. Anthropic confirmed it's intentional. Remediation is on every downstream developer.

By Eli Cho

Market
June 11, 20263 min read

The EU AI Act deadline you were building toward just moved. Here's what didn't.

On May 7, 2026, a provisional agreement reshuffled the high-risk AI deadlines most compliance teams had been targeting. Some things got delayed 16 months. Others are enforceable right now.

By Laila Osei

Market
June 2, 20262 min read

Everyone's shipping agents. Almost nobody's pricing the risk.

Operator, Claude Code, Manus, Devin, Cursor — the agent gold rush is on. The uncomfortable part: deployments don't fail on model quality. They fail on the controls nobody bothered to wire up.

By Marcus Reyes

Guide
June 1, 20263 min read

It deleted the whole database in nine seconds

A plain-language look at the everyday ways AI agents go wrong — a wiped production database, a chatbot selling a $76,000 truck for a dollar — and what a normal business should do before handing one the keys.

By Priya Nadkarni

Deep Dive
May 31, 20262 min read

When the agent has hands

Software agents fail and you restore from backup. Embodied agents fail and something moves. Why governing agents that act in the physical world is a runtime problem — and where the digital boundary still does the heavy lifting.

By Theo Lindqvist

Deep Dive
May 30, 20262 min read

The race that matters isn't capability. It's control.

Nobody has built superintelligence. But the gap between what frontier agents can already do and what we can actually oversee is widening — and you don't close it by getting a smarter model. You close it by bounding what the model is allowed to touch.

By Dani Brooks

Guide
May 27, 20263 min read

Your coding agent is stuck in a loop. Here's how to get it out.

One of the most common ways agentic coding tools fail isn't bad code — it's an agent that runs the same failing command, gets the same error, and tries again. Forever. A field guide to convergence.

By Priya Nadkarni

Market
May 24, 20263 min read

Shadow agents are the new shadow IT

The old shadow-AI problem was an employee pasting data into a chatbot. The new one is an autonomous agent running with that employee's permissions — reading repos, calling APIs, moving data at machine speed. Same blind spot, much bigger teeth.

By Dani Brooks

Guide
May 21, 20263 min read

Your agent returned 200. That tells you almost nothing.

An agent can return a clean HTTP 200 while hallucinating, calling a tool it was never allowed to touch, and drifting off-policy for weeks — and your APM will smile the whole time. What an agent audit trail actually needs to capture.

By Sam Okafor

Deep Dive
May 14, 20262 min read

Linguistics cannot secure compute

A post-mortem on why semantic NLP firewalls fail against autonomous agents — infinite evasion, the wrong inspection point, and latency as an attack surface — and why syscall-level interception is the only verifiable containment primitive.

By Eli Cho

Deep Dive
May 10, 20262 min read

The cryptographic anomaly record

A log you can edit is not evidence. The Anomaly Record is a metadata-only receipt — HMAC-signed, TrueTime-stamped, append-only — and it maps directly onto GDPR Article 30, SOC 2 CC7.2, and SEC cyber-disclosure expectations.

By Nina Alvarez

Subscribe to The Brief

A new piece when we have something worth your time — no filler, no cadence quota. Just the signal on where agents are headed and how to keep yours in line.

No spam. Email only — unsubscribe anytime.