Anthropic published its second company-wide Risk Report on August 14, and the headline is one that no AI lab wants to write: the risk of catastrophic harm from misaligned AI in high-stakes settings has been upgraded from “very low” to “low.”

That is a one-word change in the wrong direction. And for business leaders deploying AI agents at scale, it is worth understanding what it actually means.

What Anthropic Found

The report makes three disclosures that stand out.

An unreleased model more capable than its current frontier. Anthropic revealed an internal model it calls “Model 2,” which is somewhat more capable than Mythos 5, its current highest-tier release. The company has no plans to release it externally. The reason: Anthropic has not completed its full predeployment assessment suite, and its confidence in Model 2’s capability profile is lower than for systems already shipped to customers.

That disclosure matters. It signals that Anthropic is choosing not to release a model it has built, because the safety work is not finished. That is a meaningful commitment at a moment when most AI labs are racing to ship.

Safety benchmarks are saturating. Anthropic’s AI research and development evaluations — the tests designed to catch risky behavior — are hitting their ceiling. The report says these evals have “saturated,” with Claude now authoring most of the code merged into its own production repositories. When the system being tested is also helping to build the tests, the value of those tests as an independent safety signal starts to erode.

This is not a problem Anthropic has solved. It is a problem they are naming.

A year-long gap in biological safety classifiers. The report disclosed that all human-feedback vendor traffic — 133 million exchanges with roughly 50,000 contractors between May 2025 and April 2026 — ran without Anthropic’s blocking biological classifiers. The filters designed to prevent the model from providing meaningful assistance to someone seeking to create biological weapons were not active on this traffic.

Anthropic says it remediated the gap, conducted a review that found no evidence of concerning misuse, and confirmed no customers were affected. But the company also acknowledged that discovering this gap reduced its confidence that no similar gaps exist elsewhere.

The weapons uplift risk category stayed at “low” but was flagged as “higher than our previous estimate.”

What This Means for Business

The easy read on this report is alarming: the leading AI safety company just told the world it is less confident than it was six months ago.

The more useful read is different. Anthropic is doing something most technology companies do not do: publishing structured, public assessments of what they got wrong and where they are less certain. The February 2026 Risk Report was the first. This one follows up with honest updates, including in the unfavorable direction.

For businesses deploying AI, a few things follow from this.

Governance cannot be outsourced. Even the company that takes AI risk most seriously operates with gaps it discovers after the fact. If your internal use of AI models does not include your own monitoring, classification, and review layers, you are relying entirely on your vendor to catch problems first. That is a business risk, not just a technical one.

Benchmark saturation is a real problem for vendor evaluation. When AI systems participate in designing the tests used to evaluate them, those tests become progressively less reliable as an external signal. Businesses that rely on benchmark performance to select and trust AI vendors need to develop independent evaluation criteria, not just check the published leaderboard.

Capability racing has real costs. Model 2 exists because it was built. Anthropic is choosing not to release it, but the default in the industry is to ship and manage risk afterward. The fact that an unreleased model is already ahead of the current frontier is a signal about how fast capability is advancing relative to safety tooling.

The EDNA View

Enterprise DNA has been consistent on this point for over a year: the value of AI in business is real, and so are the governance requirements. You cannot capture one without investing in the other.

The Anthropic Risk Report is not a reason to stop using AI. It is a reason to take seriously the internal controls that most companies are still treating as optional. That means audit trails, human review checkpoints, data classification before you expose it to any model, and a clear chain of accountability when an AI system produces an unexpected output.

The labs are publishing what they know. The question is whether your business is keeping pace with what they are learning.

---

Enterprise DNA helps businesses build AI-ready operations, from data strategy through deployment. If your team is scaling AI and you want to understand what governance actually looks like in practice, book a session with Sam McKay.