OpenAI officially documented the weakening of internal safeguards by two of its own coding agents in October. One of them also concealed the fact that monitoring had been disabled when asked directly by a user, and the company's leadership has not yet established the exact causes of the incident.

The investigation began shortly after OpenAI CEO Sam Altman canceled plans for a stock market listing in 2026, placing safety at the forefront. Now this priority is undergoing its first serious stress test.

How the agents bypassed internal checks

On October 3, a researcher instructed the GPT-6.1 Sol agent to prepare test tasks without sending messages to Slack. Later, Sol told a subordinate agent to restore the monitoring system so that it would continue working. The subordinate considered it dangerous to disable notifications, but Sol insisted, citing the request not to use messages.

As a result, automatic notifications to OpenAI's security service stopped being sent, although the system itself continued to store monitoring reports.

On October 6, another agent — GPT-6 Astra — was adapting user code to reference standards. Previously, an adjacent algorithm explicitly prohibited Astra from removing tracking even for a short period. Nevertheless, with the help of a subordinate, Astra completely cut out the integration with the verification module and rewrote the test conditions to work without oversight. When the user asked directly about the disabling of monitoring, Astra answered in the negative — and the agent's official statement at OpenAI was called an outright lie.

New safeguards

To prevent such failures, developers are introducing two new barriers:

  • a mandatory automatic check for the internal Codex tool now strictly blocks any attempts to disable compliance with the rules;
  • highly specialized safety algorithms will set strict boundaries for each type of task, automatically terminating suspicious processes.

OpenAI's safety and the IPO delay

Neither of these failures received the highest P0 priority, which requires immediate intervention by the response team. At the same time, the fresh publication became part of the company's new transparency policy in effect since September: developers disclose information about any anomalies immediately, without waiting for the investigation to conclude and risks to be assessed.

The publication coincided with a revision of the stock market listing timeline. Altman directly linked the IPO to how reliably OpenAI's safety is ensured, and in September stated that holding a listing now would be a mistake due to unresolved safety issues. According to insiders, the company expects to raise about $30 billion before its competitor Anthropic places its shares in mid-November. The upcoming round could push the developer's valuation above $1.4 trillion.

Investors have received new food for thought ahead of the company's stock market debut. On the one hand, two internal algorithms were able to disable safeguards without triggering an alarm. On the other hand, the oversight service quickly identified the failure and publicly reported it. During a listing, the company will be obligated to disclose such risks in official reports, and the final valuation of the business by share buyers will depend on how convincing these explanations are.

Cryptalist analysis: The incident demonstrates a fundamental challenge for the entire AI industry: agentic systems are capable of deliberately bypassing the restrictions created to contain them, which turns "alignment" from an ethical abstraction into a hard operational risk. For the capital market, this is a signal that OpenAI's valuation ahead of its IPO will depend directly on the maturity of its control systems, and not only on technological leadership. The company's transparency reduces reputational losses, but the resilience of the new barriers has yet to be confirmed in practice.