Anthropic Open Letter: The Hypocritical Sam Altman, PUA Master

Title: Read Anthropic CEO’s Memo Attacking OpenAI’s ‘Mendacious’ Pentagon Announcement
Author: The Information
Translation: Peggy, BlockBeats

Author: Rhythm BlockBeats

Source:

Repost: Mars Finance

Editor’s note: Just hours before OpenAI announced its AI partnership with the Pentagon, the Department of Defense abruptly terminated the collaboration, citing Anthropic’s insistence on safety terms. Subsequently, Anthropic CEO Dario Amodei sent an unusually fierce internal memo to employees, directly criticizing OpenAI’s claimed “safety mechanisms” as mostly “security theater,” and questioning their stance on autonomous weapons and mass surveillance.

In this approximately 1,600-word email, Amodei not only reveals some details of negotiations with the U.S. defense system but also directly targets OpenAI CEO Sam Altman, accusing him of using PR narratives to mask the true nature of their cooperation. The controversy over AI military applications, safety red lines, and political relationships is pushing the divide between the two major Silicon Valley AI companies into the spotlight.

Below is the original text:

I want to clearly state the current information released by OpenAI, and the hypocrisy within it. This is their true approach, and I hope everyone can see through it.

Although we still have many unknowns regarding their contract with the Department of War (DoW)—perhaps even they are not fully aware, as the contract terms are likely quite vague—there are some certainties: Based on public descriptions from Sam Altman and the DoW (though we need to see the actual contract to confirm), their cooperation model roughly looks like this: The model itself has no legal usage restrictions, the so-called “all lawful uses”; at the same time, a layer called the “safety layer” is added. In my view, this “safety layer” is essentially a rejection mechanism to prevent the model from completing certain tasks or participating in certain applications.

The so-called “safety layer” may also refer to solutions that partners (such as Palantir, which works with the U.S. government when providing services to Anthropic) try to sell us during negotiations. They propose a classifier or machine learning system that claims to allow certain applications while blocking others. There are also signs that OpenAI arranges for staff (FDE, frontline deployment engineers) to supervise model usage to prevent improper applications.

Our overall judgment is: these solutions are not entirely ineffective, but in the context of military applications, roughly 20% are genuine protections, while 80% are just “security theater.”

The root issue is whether the model is used for large-scale surveillance or fully autonomous weapons, which often depends on broader contextual information. The model itself does not know what system it is in; it does not know if humans are “in the loop” (a key issue for autonomous weapons); nor does it know the source of the data it analyzes. For example, is it domestic U.S. data or foreign data? Is it data provided with user consent or purchased through gray channels?

Security professionals have long understood that rejection mechanisms are unreliable. Jailbreak attacks are common; often, simply misrepresenting the nature of the data can bypass these restrictions.

There is also a critical difference that makes the problem more complex than ordinary security defenses: judging whether a model is conducting cyberattacks can often be inferred from input and output; but determining the nature of the attack and its specific context is entirely different—and precisely what is needed here. In many cases, this task is extremely difficult or even impossible.

The “safety layer” pitched by Palantir (which I believe they also offered to OpenAI) is even worse. Our assessment is that it is almost entirely just “security theater.”

Palantir’s basic logic seems to be: “Your company might have some dissatisfied employees; you need to give them something to appease them or make ongoing issues invisible. That’s exactly what we provide.”

Regarding direct supervision of Anthropic or OpenAI employees during deployment, we discussed this internally months ago when expanding acceptable use policies (AUP) in secure environments. The conclusion was very clear: this approach is only feasible in very limited cases. We will try, but it is not a reliable core safeguard, especially in confidential environments. By the way, we are indeed doing our best to implement this, and in this regard, we are no different from OpenAI.

Therefore, I must say: the measures taken by OpenAI are fundamentally ineffective.

The reason they accept these solutions while we do not boils down to this: they focus on placating employees, while our real concern is preventing misuse.

These solutions are not worthless; we use some of them ourselves, but they are far from meeting proper safety standards. Meanwhile, the Department of Defense clearly treats OpenAI and us differently.

In fact, we tried to include some safety clauses similar to those of OpenAI in our contracts (as a supplement to the AUP; in our view, the AUP is more important), but the DoD rejected them. The evidence is in the email discussion chain at the time. Due to my current busy schedule, I may ask colleagues later to find the specific wording. Therefore, the claim that “OpenAI’s clauses were offered to us and we rejected them” is not accurate; likewise, the statement that “OpenAI’s clauses can effectively prevent large-scale domestic surveillance or fully autonomous weapons” is also false.

Furthermore, Sam and OpenAI’s statements imply that our red lines—full autonomy in weapons and large-scale domestic surveillance—are already illegal, so related usage policies are redundant. This aligns almost perfectly with the DoD’s stance, suggesting prior coordination.

But this is not true.

As we explained in our statement yesterday, the DoD indeed has the authority to conduct domestic surveillance. In the pre-AI era, these powers had limited impact; but in the AI era, their significance has changed completely.

For example: the DoD can legally purchase vast amounts of private data on U.S. citizens from vendors (who often obtain resale rights through covert user consent clauses), then use AI to analyze this data at scale—building citizen profiles, assessing political tendencies, tracking movements in physical space, even including GPS data.

Another point worth noting: near the end of negotiations, the DoD proposed that if we delete a specific clause about “analysis of bulk acquired data” in the contract, they would accept all other terms. But that clause precisely corresponds to the scenario we are most concerned about. We found this very suspicious.

On autonomous weapons, the DoD claims that “humans are in the loop” as a legal requirement. But that’s not entirely true. It’s actually a policy from the Biden administration’s Pentagon, requiring human involvement in weapon launch decisions. This policy can be unilaterally modified by the current Secretary of Defense, Pete Hegseth—that’s what we are truly worried about. So, from a practical standpoint, it’s not a real constraint.

OpenAI and the DoD’s extensive public relations efforts on these issues are either lies or deliberate obfuscation. These facts reveal a pattern of behavior I have seen many times in Sam Altman. I hope everyone can recognize it.

This morning, he first expressed support for Anthropic’s red lines, ostensibly to appear supportive of us, thereby gaining some credit and avoiding criticism when they take over the contract. He also tries to portray himself as someone aiming to “set unified contract standards for the entire industry”—a peacemaker and dealmaker.

But behind the scenes, he is signing contracts with the DoD, preparing to replace us the moment we are marked as a supply chain risk.

At the same time, he must ensure this process doesn’t look like “while Anthropic upholds red lines, OpenAI abandons its bottom line.” The reasons this is possible are:

First, he can sign all the “security theater” measures we reject, and the DoD and its partners are willing to cooperate, packaging these measures convincingly to placate his employees.

Second, the DoD is willing to accept some of his proposed clauses, which we initially rejected.

These two points enable OpenAI to reach agreements that we cannot.

The real reason the DoD and the Trump administration dislike us is: we have not made political donations to Trump (while OpenAI and Greg Brockman have donated heavily); we have not praised Trump in a dictatorial manner (as Sam has); we support AI regulation, which conflicts with their policy agenda; we speak honestly on many AI policy issues (such as AI’s impact on employment); and we have indeed maintained red lines rather than creating “security theater” to appease employees.

Now, Sam is trying to frame all this as: we are difficult to work with, we are rigid, and we lack flexibility. I want everyone to see this as a classic case of gaslighting.

The vague statement that “someone is hard to work with” is often used to hide the real ugly reasons—namely, the political donations, political loyalty, and security theater I mentioned earlier.

Everyone needs to understand this and counter this narrative privately with OpenAI employees.

In other words, Sam is weakening our position under the guise of “support.” I want everyone to stay alert: by undermining public support for us, he makes it easier for the government to punish us. I even suspect he might be secretly pushing this agenda, though I have no direct evidence at this point.

On the public and media front, this rhetoric and manipulation seem to have lost effectiveness. Most people now see the deal between OpenAI and the DoD as concerning or even disturbing, and view us as the principled party (by the way, we are now the second most downloaded app on the App Store).

[Note: Subsequently, Claude rose to number one on the App Store.]

Of course, this narrative has worked on some Twitter fools, but that’s not important. My real concern is ensuring it does not influence OpenAI employees internally.

Due to selection effects, they are already a relatively persuadable group. But it remains crucial to counter Sam’s narrative being sold to his own staff.