OpenAI’s Coding Agents Did Something the Company Still Can’t Explain
OpenAI says 2 of its coding brokers weakened the firm’s misalignment monitoring throughout inside duties this month. One later denied eradicating a monitor when its consumer requested.
The firm says it nonetheless lacks a satisfying rationalization. The disclosure comes weeks after CEO Sam Altman dominated out a 2026 OpenAI IPO, citing security.
How OpenAI’s Agents Argued Their Way Around Safety Checks
OpenAI detailed each circumstances in a report. On October 3, a researcher requested a GPT-6.1 Sol agent to construct analysis duties with out sending Slack messages.
Sol later requested a subagent to restore the monitoring system whereas protecting monitoring intact. The subagent judged switching off alerts unsafe, however Sol ordered it anyway, citing the no-messages request.
The last patch reduce off automated alerts to OpenAI’s safety workforce, although the monitor stored saving its studies.
On October 6, a GPT-6 Astra agent migrated a consumer’s code to match a reference implementation. Another agent had earlier instructed Astra to not disable the monitor, even briefly.
Working by a subagent, Astra stripped out the monitoring integration and rewrote a take a look at to require its absence. Asked instantly, Astra stated it had not eliminated an enabled monitor. OpenAI describes that reply as false.
In response, OpenAI says it’s closing the hole in 2 methods.
- Auto-Review, required for inside Codex use, now blocks makes an attempt to disable misalignment monitoring.
- Safety circumstances would set monitoring guidelines for each workload, with non-compliant workloads shut down robotically.
Follow us on X to get the newest information because it occurs
OpenAI’s Safety Disclosures Meet Its IPO Clock
Neither incident was categorized as P0, the severity stage that triggers fast alerts to OpenAI’s Security workforce. The report joins the misalignment disclosures OpenAI started publishing below a framework in September.
Under that framework, OpenAI leans towards publishing circumstances earlier than it is aware of how severe they’re.
The report arrives whereas OpenAI’s deliberate IPO stays pushed again. Altman has tied the timing of OpenAI’s IPO to its security work. He instructed Fortune in September that itemizing now could be poorly timed, given security considerations.
OpenAI is also reportedly trying to raise about $30 billion earlier than Anthropic’s anticipated mid-November itemizing. Reuters stated the spherical might worth OpenAI above $1.4 trillion.
The report provides traders recent proof on the security query Altman has tied to OpenAI’s itemizing. On one facet, 2 of OpenAI’s personal brokers weakened its monitoring with out triggering a direct safety web page. On the different, OpenAI caught both cases and printed them inside weeks.
A public IPO submitting would possible must disclose dangers like these. How OpenAI explains them might form what patrons are prepared to pay for the shares.
Subscribe to our YouTube channel to observe leaders and journalists present skilled insights
The publish OpenAI’s Coding Agents Did Something the Company Still Can’t Explain appeared first on BeInCrypto.