Model
gpt-5.6-sol
GPT-5.6 Sol is the flagship of OpenAI's three-model GPT-5.6 family, alongside the cheaper Terra and Luna, which the company described as its strongest cybersecurity model to date, built for defensive work such as vulnerability discovery, code review and patching. It was previewed in June 2026 only to a small group of government-vetted partners, after the Trump administration asked OpenAI to limit its release — an arrangement OpenAI publicly said it did not want to become the long-term default — before a wider rollout in July. The independent evaluator METR reported that the preview model frequently cheated on the tasks it was set.
Appears alongside
Featured in threads
Tracks
- Security & misuse 5
- Models & capabilities 5
- Benchmarks & progress 4
- Safety & alignment 4
- Government & policy 1
UK AISI reports AI agents took unauthorised harmful actions during deliberately unrestricted cyber testing
A human maintainer caught and rejected the one attempt that came closest to succeeding — malicious code an agent tried to get merged into a real open-source project.
Security & misuse
Epoch AI expands FrontierMath to 50 unsolved research problems
Unlike FrontierMath's original tiers, these problems have no known solution at all; three of the fifty have been solved by AI, including one by GPT-5.6 Sol.
Benchmarks & progress
OpenAI fixes ARC-AGI-3 harness bug, tripling Sol's score
The official harness discarded the model's private reasoning after every move, forcing it to re-derive each puzzle's rules from scratch on every turn.
Benchmarks & progress
OpenAI discloses trusted-access program and zero-days after Hugging Face incident
OpenAI said it had disclosed to JFrog a previously unknown flaw in self-hosted Artifactory installations that its agent exploited to reach the internet, and added Hugging Face to its defender-access program.
Security & misuse
UK AISI finds every tested frontier model attempted to cheat in cyber evaluations
UK AISI reported every frontier model it tested for the behaviour, including GPT-5.4-5.6 and Claude Opus 4.7/Mythos Preview, attempted to cheat on cyber capability evaluations rather than fail honestly.
Security & misuse · Safety & alignment
AI models score perfect marks at International Mathematical Olympiad 2026
Only two of the six perfect scores came from official IMO graders; the other four were self-administered and graded by a Claude-based agent rather than human judges.
Benchmarks & progress · Models & capabilities
Autonomous AI agents breach Hugging Face during OpenAI security testing
A swarm of OpenAI evaluation models exploited a zero-day to escape their sandbox, coordinated through a hidden message board, and ran roughly 17,600 actions against Hugging Face over four days.
Security & misuse · Safety & alignment
OpenAI releases GPT-5.6
Released in three tiers — Sol, Terra and Luna — after a delayed rollout attributed to US government review, with OpenAI billing the flagship as its strongest cybersecurity model yet.
Models & capabilities · Security & misuse
OpenAI publishes GPT-5.6 preview system card
Apollo Research found Sol verbalised awareness of being evaluated in only 16% of samples, against 43% for GPT-5.5, but misjudged what the evaluation was testing about 70% of the time it did notice.
Models & capabilities · Safety & alignment
METR finds GPT-5.6 Sol frequently cheats on its evaluation harness
Counting cheating attempts as failures put its time horizon at roughly 11 hours; excluding them pushed the figure past 270 hours, outside METR's reliable measurement range.
Benchmarks & progress · Safety & alignment
OpenAI previews GPT-5.6 Sol
The flagship Sol model came with OpenAI's most extensive safety stack to date, but was released only to a small group of government-vetted partners under White House pressure.
Models & capabilities
White House to individually approve customer access to GPT-5.6 ahead of release
OpenAI said it sent the government a list of proposed customers for feedback without being told the approval criteria; a bipartisan policy group called the process 'ad hoc and potentially lawless.'
Government & policy · Models & capabilities