Three former OpenAI safety researchers pushing back against their recent terminations published an open letter denying claims that they improperly shared confidential company information with external safety entities. Researchers Jasmine Wang, Jascha Sohl-Dickstein, and Misha Khalman pushed back against OpenAI’s official narrative, warning that firing key safety personnel creates a dangerous chilling effect across the company’s internal research culture. The trio previously served on critical internal safety bodies, including OpenAI’s Safety and Security Committee, Safety Advisory Group, and Misses Advisory Council.
OpenAI defended the firings by claiming the researchers engaged in willful misconduct by transferring internal documents to an outside safety organization. Management insisted the dismissals were necessary to safeguard confidential intellectual property. However, the ousted researchers countered that their interactions with external safety groups were essential for establishing independent evaluation standards for frontier artificial intelligence models. They noted that AI safety research requires open communication with outside experts to identify vulnerabilities, evaluate systemic risks, and build unbiased safety frameworks.
In their joint statement, the researchers expressed concern over how internal safety disputes get handled inside top AI labs. They stressed that punishing employees for consulting outside experts creates an environment where staff members fear raising safety concerns. When researchers worry about losing their jobs for raising red flags or seeking second opinions on model behavior, critical safety reviews get suppressed. They argued that OpenAI’s heavy-handed response undermines the open safety culture the company previously claimed to champion.
The dispute highlights a growing rift between commercial product teams and safety evaluation groups inside frontier labs. Reporting indicates OpenAI fired the researchers after discovering they shared technical research modules and model evaluation frameworks with external testing groups. The shared materials focused on assessing model capabilities, detecting autonomous agent failures, and testing alignment guardrails. While OpenAI viewed the file transfers as policy violations, the researchers argued their actions aligned directly with industry best practices for independent safety verification.
OpenAI’s official stance drew immediate pushback from internal research staff and external AI governance advocates. Critics pointed out that relying solely on internal safety audits creates clear conflicts of interest when commercial release deadlines loom. Independent evaluations provide essential checks against rushed deployments, ensuring models meet strict safety benchmarks before hitting public markets. Terminating researchers for collaborating with outside auditors signals that commercial speed takes priority over independent oversight.
The letter also addressed specific personal allegations brought against individual team members during the internal investigation. Wang clarified details regarding her termination, explaining that OpenAI initially claimed she lost access due to recruitment conflicts before shifting to data security claims. She noted that accessing internal tools via personal devices was standard practice across research teams and had never drawn disciplinary action previously. The researchers maintained that OpenAI used security policies selectively to justify removing vocal safety advocates.
As artificial intelligence models grow more autonomous and capable, maintaining open, transparent safety pipelines remains essential. Suppressing internal debate and punishing collaboration with outside safety organizations puts public trust at risk. The ousted researchers urged OpenAI to honor its founding commitments to public safety, restore transparent review channels, and protect staff members who raise safety concerns from retaliation. Creating safer AI systems requires protecting the voices tasked with keeping development on track.