Anthropic is having a month

In the world of technology, where trust is a currency as valuable as computing power, Anthropic is currently going through one of the most difficult periods in its history. The company, which since its inception has positioned itself as a "safer and more responsible alternative to OpenAI," has experienced two spectacular stumbles resulting from mundane human errors in just seven days. It is a painful reminder that even the most advanced security systems and AI algorithms are ultimately dependent on a human who might simply "forget to check the right box."

The situation is particularly ironic because Anthropic builds its image around rigorous research into risk and ethics. The startup, founded by the Amodei siblings, employs leading researchers in the field of AI safety and regularly publishes extensive reports regarding the threats posed by the development of large-scale models. Currently, the company finds itself at the very center of public debate, engaging in disputes with institutions such as the Department of Defense, which only intensifies the weight of the recent incidents. When you preach the need for global control over powerful technology, publicizing confidential data becomes more than just a PR blunder – it becomes an undermining of the foundations of your mission.

Leak of three thousand files and a premature premiere

The first alarm signal appeared when Fortune reported a massive leak of internal documentation. Due to a configuration error, nearly 3,000 internal files became publicly available to anyone who knew where to look. The scale of this incident is staggering, considering that Anthropic operates on the most closely guarded secrets of the tech industry. Among the documents that saw the light of day were not only routine notes but also draft blog posts describing a completely new, powerful AI model that the company had not yet officially announced.

Disclosing information about a new model ahead of time is a logistical and strategic nightmare for a technology company. In an industry where the state-of-the-art changes every few weeks, control over the narrative and the timing of breakthrough announcements is crucial for maintaining a competitive advantage. The fact that such sensitive data leaked "by accident" calls into question the internal information flow procedures in a company that wants to teach the world how to safely handle data and artificial intelligence.

The "one click" error and consequences for image

Just a few days after the file incident, another event occurred on Tuesday. According to available information, an Anthropic employee failed again in the simplest possible aspect – verifying access settings. "Forgetting to check a box" sounds like a trivial mistake in an accounting office, but in the context of a company operating with billions of dollars and strategically significant technology, it is an error of colossal importance. The recurrence of these events in such a short interval suggests a systemic problem with operational security culture, which contrasts with their theoretical approach to algorithmic safety.

The problem is that Anthropic is not perceived as an ordinary startup. Thanks to partnerships with giants like Amazon and Google, this company is a foundation of the new technological infrastructure. Their Claude model is used by corporations to process sensitive business data. If the creators of the tool cannot keep track of their own draft blog posts or internal databases, corporate clients may start asking uncomfortable questions about the security of their own assets entrusted to the company's systems.

Between theory and practice of security

There is a clear disconnect between what Anthropic calls AI Safety (model security, avoiding hallucinations, lack of bias) and traditional Cybersecurity (data protection, access control). One might get the impression that the company focused so much on preventing a hypothetical "machine rebellion" and the existential risks associated with AGI that it neglected the basic principles of digital hygiene that apply to every company in the IT sector. It is a classic case of "the shoemaker's children going barefoot" – except in this case, trade secrets and investor trust are at stake.

• Loss of control over the product roadmap: The leak of a draft about a new model forces the company into reactive mode instead of a planned marketing offensive.
• Erosion of institutional trust: Disputes with government bodies become harder to win when opponents can point to real negligence in data protection.
• Internal pressure: Two human errors in a week is a signal to the board that the pace of development may be exceeding the team's operational capabilities.

Analyzing these events from an editorial perspective, it is clear that Anthropic fell victim to its own success and the pace at which it tries to chase the competition. Building "safe AI" requires time, calm, and rigor, while the race with OpenAI forces haste. These two incidents show that the greatest threat to artificial intelligence companies is not malicious algorithms or outside hackers, but the fatigue and inattention of employees who skip procedures in the daily rush. In the coming months, Anthropic will have to prove that it can manage not only technological risk but, above all, human risk, which in practice turns out to be a much more difficult task.