OpenAI says GPT-5.5 matches Mythos Preview in latest cybersecurity tests
Key Points
- OpenAI says GPT-5.5 performed on par with Mythos Preview in cybersecurity tests.
- CEO Sam Altman criticized the hype around Mythos as fear‑based marketing.
- Trusted Access for Cyber pilot, launched in February, governs who can test frontier models.
- GPT-5.4-Cyber was released to vetted users via the trusted‑access list last month.
- Initial GPT-5.5-Cyber rollout will be limited to critical cyber defenders within days.
- Altman warned future models deemed dangerous will also see restricted releases.
OpenAI announced that its upcoming GPT-5.5 model performed on par with the heavily promoted Mythos Preview in recent cybersecurity evaluations. CEO Sam Altman criticized the hype surrounding Mythos, calling it fear‑based marketing, while reiterating that the new model will initially be available only to a select group of critical cyber defenders. The company’s Trusted Access for Cyber pilot, launched in February, continues to serve as the gateway for researchers and enterprises to test frontier models under strict safeguards.
OpenAI revealed that its forthcoming GPT-5.5 model achieved results comparable to the much‑talked‑about Mythos Preview in a series of cybersecurity tests. The findings suggest that Mythos’ apparent breakthrough was not unique to a single model but rather reflected broader gains in long‑horizon autonomy, reasoning and coding, according to the Artificial Intelligence Security Institute.
During a recent interview on the Core Memory podcast, OpenAI chief executive Sam Altman took aim at what he called “fear‑based marketing” surrounding limited‑release AI systems. While he conceded that Mythos was likely a solid model for cybersecurity, Altman likened the hype to “selling a bomb shelter for $100 million.” He warned that similar rhetoric would accompany future releases of models deemed too dangerous for the public.
OpenAI’s Trusted Access for Cyber pilot, introduced in February, lets vetted security researchers and enterprise teams verify their identities and express interest in studying the company’s frontier models for legitimate defensive work. The program underpins the controlled rollout of GPT‑5.4‑Cyber, a variant fine‑tuned for enhanced cyber capabilities and fewer usage restrictions.
Last month, OpenAI announced that it was using the trusted‑access list to manage the limited launch of GPT‑5.4‑Cyber. The company now says the initial release of GPT‑5.5‑Cyber will follow a similar path, reaching “critical cyber defenders in the next few days.” The approach reflects OpenAI’s strategy of staggered, highly restricted deployments to mitigate misuse while still providing powerful tools to those defending against threats.
Altman’s comments highlight a tension within the AI community: balancing the need to curb malicious exploitation of advanced models against the desire to empower security professionals with cutting‑edge capabilities. By restricting access to a curated pool of defenders, OpenAI hopes to demonstrate that robust safeguards can coexist with rapid innovation.
Industry observers note that the alignment between GPT‑5.5 and Mythos may signal a shift away from model‑specific hype toward a focus on underlying improvements that benefit a range of applications. If the performance gap truly narrows, future marketing could emphasize the broader progress in AI reasoning rather than isolated breakthroughs.
OpenAI has not disclosed a public release date for GPT‑5.5‑Cyber beyond the near‑term rollout to trusted users. The company continues to refine its access controls, expecting that lessons learned from the pilot will inform wider deployments once confidence in safety mechanisms grows.