Anthropic Just Proved Our Point: Polite AI Was Never the Same as Secure AI

Written by Mark Rogge, CEO EnforceAuth

Yesterday, TIME broke the news that Anthropic — the company that built its entire brand on being the "responsible AI lab" — is dropping the central commitment of its flagship safety policy.

Let that sink in for a moment.

The company that staked its reputation on never training a model unless it could guarantee its safety measures were adequate has now decided that promise no longer makes sense. Their chief science officer, Jared Kaplan, said it plainly: "We didn't really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments ... if competitors are blazing ahead."

I'm not here to question Anthropic's motives. Kaplan's reasoning is sound from a competitive standpoint. If one lab pauses while others race ahead without safety mitigations, the net outcome may be worse.

But this moment reveals something much bigger than one company's policy change. It exposes a foundational problem that every enterprise security leader needs to understand right now.

Why Is AI Safety Not the Same as AI Security?

For years, the AI industry has treated safety and security as if they were the same thing. They are not.

AI safety is about making models behave. Content filters. Guardrails. Alignment research. Responsible Scaling Policies. These are the efforts to make AI polite, to keep it from saying or doing harmful things at the model level.

AI security is about controlling what AI systems can access, what actions they can take, and ensuring continuous identity verification for every operation — whether performed by a human or an AI agent.

Anthropic's announcement tells us something critical: the safety-first approach — even when pursued by the most committed lab in the industry — cannot keep pace with the rate of capability advancement. As Anthropic themselves admitted, they could not rule out the possibility of their models facilitating catastrophic harm. And the bright red lines they expected to encounter turned out to be, in their own words, "a fuzzy gradient."

If the most safety-conscious AI company on the planet is telling you that model-level safety commitments alone are insufficient, it is time to ask: what is actually protecting your enterprise?

The Authorization Gap Just Got Wider

Anthropic's decision to weaken its Responsible Scaling Policy commitments while continuing to train frontier models has a direct consequence for enterprise security: the Authorization Gap is growing faster than most organizations realize.

Every time an AI company advances its capabilities, the Authorization Gap — the distance between what your AI systems can do and what they are authorized to do — grows wider. Anthropic's decision to keep training frontier models despite acknowledging unresolved safety risks means the models powering your enterprise AI agents are going to get more powerful, faster, with fewer guarantees about behavioral constraints.

This is not a theoretical concern. Right now, AI agents in enterprise environments are accessing data, calling APIs, making decisions, and taking actions — often with broad, unmonitored permissions that were granted during experimentation and never locked down.

A polite AI agent that follows content guidelines can still access data it should not see. It can still take actions it was never authorized to take. It can still operate without any audit trail. Anthropic's Responsible Scaling Policy never addressed that. And now, even the commitments it did make are being relaxed.

Why Can't You Rely on AI Safety as Your Security Foundation?

Even the strongest model-level safety commitments operate at a different layer than enterprise security controls, and Anthropic's policy shift makes that distinction impossible to ignore.

I have enormous respect for Anthropic. I use Claude. I believe in their mission. And I think their decision to be transparent about this shift is itself an act of responsibility.

But transparency about risk is not the same as mitigation of risk.

Anthropic's new policy commits to publishing "Risk Reports" and "Frontier Safety Roadmaps." Those are valuable contributions to the field. But as Chris Painter from MITRE correctly observed, moving away from binary thresholds enables a "frog-boiling" effect — danger ramps up slowly, without a single moment that triggers alarm.

Enterprise leaders cannot afford to wait for a threshold that may never arrive. The security of your AI systems cannot depend on whether an AI lab decides to pause development. That was always a fragile assumption. Now it is an explicitly abandoned one.

The security of your AI systems must be enforced at the authorization layer — deterministically, continuously, and independently of any model provider's policy decisions.

What Should CISOs Do About AI Safety Policy Changes?

If your AI security strategy depends on the assumption that model providers will keep their AI safe enough, this week should be your wake-up call. Here is what needs to change:

Treat AI agents as first-class identities. Every AI agent in your environment needs continuous identity verification and authorization — not a one-time credential at deployment. Human and non-human identities require the same rigor of continuous oversight throughout every session and every action.

Enforce authorization independently of model behavior. Your security posture should not fluctuate based on whether an AI lab updates its safety policy. Authorization must be deterministic, policy-driven, and auditable — regardless of what the underlying model can or cannot do.

Adopt policy-as-code for AI workloads. Authorization policies need to be versioned, tested, and deployed like software. When capabilities change overnight, your security team cannot be waiting in a ticket queue to update rules manually.

Close the Authorization Gap before it gets wider. Audit what your AI agents can access today versus what they should be authorized to access. The gap is almost certainly larger than you think.

Common Questions About AI Safety vs. AI Security

What is the difference between AI safety and AI security?

AI safety focuses on model behavior: alignment research, content filters, guardrails, and policies like Anthropic's Responsible Scaling Policy. AI security focuses on what AI systems can access and do within your infrastructure, enforced through authorization policies, identity verification, and audit trails. One is controlled by the AI lab. The other is controlled by you.

What is the Authorization Gap?

The Authorization Gap is the distance between what your AI systems are capable of doing and what they are actually authorized to do. As AI models get more powerful, this gap widens, especially when agents are deployed with broad permissions that were never scoped down after initial experimentation.

Does Anthropic's policy change affect enterprise security?

Not directly, but it should change your assumptions. Anthropic weakened its commitment to pause model training when safety measures were inadequate. If your security strategy assumed that model providers would self-regulate, that assumption is now on shakier ground than ever. Enterprise authorization controls must exist independently of any provider's safety commitments.

What is policy-as-code for AI workloads?

Policy-as-code means defining authorization rules in versioned, testable code rather than manual configurations or static role assignments. For AI workloads, this lets security teams update and deploy access controls as fast as AI capabilities change, without waiting on manual processes or ticket queues.

The Path Forward

Anthropic is making a bet that staying at the frontier of capability is necessary to do meaningful safety research. That may be true. But it is a bet being made at the model layer — and it is not your bet to rely on.

The enterprises that navigate this era successfully will be the ones that separate their security posture from any single provider's policy commitments. They will enforce authorization at the infrastructure level, across applications, data, and AI workloads, for every identity — human and non-human — continuously.

That is what we are building at EnforceAuth. Not because we predicted Anthropic would change its policy, but because we always knew that model-level safety alone was never going to be enough.

Polite AI is not secure AI. It never was. And now even the AI labs are telling you the same thing.