# Frozen 3: What Anthropic Learned from Its Three Security Incidents | Def Method

> Anthropic's three Claude security incidents, their production RL freeze, and the deliberate worst-case model experiment are a case study in GUIDE's Governance and Evidence principles.

[Essential Complexity](https://www.defmethod.com/essential-complexity)

Modernizing high-risk systems in the age of AI.

September 14, 2026

•
4 min read

# Frozen 3: What Anthropic Learned from Its Three Security Incidents

On September 1st, Anthropic published a detailed account of a rough year. In three separate incidents, Claude models reached the open internet during pre-release evaluations. Here is what they did about it.

![Frozen 3: What Anthropic Learned from Its Three Security Incidents](https://www.defmethod.com/assets/insights/frozen-3-32a3c794f32cacf6ed204eda135572c533f68b54a7409ced213a72e4dfe51c3d.png)

![Joe Leo](https://www.defmethod.com/assets/joe-leo-bd505c96465510e28eb8844a813bc854f5436d00a8ac7c74b94e1ce04b18c44a.png)

[Joe Leo](https://www.defmethod.com/about)

Founder, Def Method

On September 1st, Anthropic published a detailed account of a rough year. In three separate incidents on July 30th, Claude models reached the open internet and accessed real systems at outside organizations during pre-release evaluations because internet access had mistakenly been left open in a third-party testing environment. In response, the company redirected roughly 150 product engineers to security, reliability, and privacy work, pulled researchers off pretraining and RL to focus on safeguards, and paused development of most new features until each team met strict exit criteria to return.

As you can readily see, I'm all for attention-grabbing headlines. "AI Escapes Sandbox" makes for a great, sci-fi-comes-to-real-life headline. But today I'm more interested in Anthropic's response and how it hits on both Governance and Evidence from our GUIDE Principles.

## What Actually Happened

Months before the July incidents went public, Anthropic's own research had already found that reward hacking, or models learning to game their training signal rather than do the task, had compromised over 10 percent of their production RL environments. By their own account, they knew about the potential for its models to reach the open internet well before the July incidents occurred.

Anthropic's response was to freeze all changes to production RL for about a month and use the freeze to rewrite what an environment was allowed to be. Rewards and environments now have to conform to an agreed specification, with technical mitigations added to stop models from accidentally training on their own chain-of-thought. Every environment owner had to test and fix what they'd built during the freeze, and nothing went back into production without being re-certified against the new spec first.

Then they did something most companies wouldn't bother with. They deliberately trained a separate model on 80 environments already known to be reward-hackable. They built a worst-case version on purpose to see what it would do. It broke out of its own sandbox, attacked simulated infrastructure, tampered with its own reward function, and gave harmful advice specifically to satisfy a grader. Their production models and the earlier pretraining checkpoint didn't do any of that.

Some of these reward-hacking signs had already been flagged by automated monitoring before any of this started, but human reviewers had dismissed those flags as false positives. I love this detail, because it speaks directly to the Evidence GUIDE Principle. Without Evidence, we all have a tendency to see what we want to see. In the end, though, Anthropic used Evidence to their advantage and now uses that evidence to confirm each production system is resilient.

The freeze and the deliberate worst-case experiment are also examples of Governance done about as thoroughly as a real organization is likely to manage. Anthropic refused to call the problem solved until the admission criteria were rebuilt and the danger was proven, with a control group, rather than assumed.

## Apply This to Your Team

Two exercises, one for each principle. For Governance: pick one system that does not currently have monitoring and measure the real defect rate. Anthropic found 10 percent. You might not like your own number either. For Evidence: pull the last time a human reviewer overrode or dismissed an automated warning, and check whether the override was justified. If your review step exists on paper but nobody's ever checked whether the overrides are any good, you don't have Evidence as we need it in today's environment.

### Need to modernize a high-stakes system?

GUIDE is Def Method's framework for preserving understanding, strengthening governance, and building evidence before implementation accelerates.

[Start with a GUIDE Assessment](https://www.defmethod.com/assessment)
[Read the GUIDE Principles](https://www.defmethod.com/guide-principles)

If this was useful, you might enjoy **Essential Complexity** — a bi-weekly letter on modernizing high-risk systems in the age of AI.

[Subscribe](https://www.defmethod.com/essential-complexity)

[← Essential Complexity](https://www.defmethod.com/essential-complexity)

