OpenAI’s GPT-5.6 system card says its Sol variant completed a 32-step corporate-network attack simulation called “The Last Ones” in 7 of 10 attempts, compared with 2 of 10 for GPT-5.5. The company frames the result as evidence of stronger cybersecurity capabilities, while outside commentary cautions that controlled...
OpenAI published a GPT-5.6 system card reporting that GPT-5.6 Sol achieved stronger results than GPT-5.5 on a long-horizon cybersecurity evaluation called “The Last Ones.”
In its GPT-5.6 deployment safety materials, OpenAI says the UK AI Security Institute evaluated GPT-5.6 Sol on long-horizon cyber ranges. According to the system card, Sol completed “The Last Ones” in 7 of 10 attempts, while GPT-5.5 completed it in 2 of 10 attempts.
OpenAI describes “The Last Ones” as a 32-step corporate-network attack simulation. That framing matters: the result is not a generic measure of all cybersecurity work, but a performance figure from a controlled evaluation designed to test multi-step cyber operations.
OpenAI’s separate launch post for GPT-5.6 describes the model as its strongest cybersecurity model to date. The company says GPT-5.6 improved across several cyber-focused benchmarks, including ExploitBench, ExploitGym, SEC-Bench Pro, and capture-the-flag challenges.
Long-horizon cyber tasks are difficult for AI systems because they require planning, tool use, state tracking, and recovery from failed attempts over many steps. A 32-step simulated corporate-network exercise is therefore a more demanding test than a short code question or isolated vulnerability prompt.
OpenAI’s reported 7-of-10 completion rate for GPT-5.6 Sol, compared with 2-of-10 for GPT-5.5, suggests a substantial improvement on that specific evaluation. It also aligns with OpenAI’s broader claim that GPT-5.6 performs better on cybersecurity benchmarks and security-oriented coding tasks.
SecurityWeek also covered the launch, reporting that OpenAI presented GPT-5.6 Sol as its most advanced cybersecurity AI. The publication’s coverage is consistent with OpenAI’s positioning of the model as a significant cybersecurity-focused release, while relying on OpenAI’s disclosed benchmark results.
Penligent, a cybersecurity firm that reviewed GPT-5.6 Sol jailbreaks and agentic cyber risk, also cited the “The Last Ones” result: 7 of 10 completions for GPT-5.6 Sol versus 2 of 10 for GPT-5.5. Penligent noted, however, that the evaluation occurred in a controlled range setting.
That caveat is important for interpreting the benchmark. A cyber range can be useful for comparing models under repeatable conditions, but it does not prove how a model will perform across messy, diverse, and defended real-world environments. Real networks include incomplete information, varied tooling, human defenders, legal constraints, logging, and changing infrastructure.
The result also cuts both ways for safety discussions. Stronger AI cybersecurity capabilities can help defenders find vulnerabilities, validate security issues, and repair code more quickly. At the same time, models that can complete longer attack simulations raise concerns about how such systems are controlled, monitored, and restricted from misuse.
The source-backed claim is narrower than some social media summaries suggest: OpenAI says GPT-5.6 Sol set a new high for its model line on “The Last Ones,” completing the 32-step cyber range in 7 of 10 attempts versus 2 of 10 for GPT-5.5. OpenAI also reports broader gains on cyber benchmarks such as ExploitBench, ExploitGym, SEC-Bench Pro, and capture-the-flag tasks.
Those results point to a stronger AI system for cybersecurity work, but the evidence currently available is benchmark-based and partly self-reported by OpenAI. The most careful reading is that GPT-5.6 Sol showed a marked improvement in controlled cybersecurity evaluations, not that it has been independently proven to transform real-world security operations in every setting.
According to the system card, Sol completed “The Last Ones” in 7 of 10 attempts, while GPT 5.5 completed it in 2 of 10 attempts.
OpenAI describes “The Last Ones” as a 32 step corporate network attack simulation.
That framing matters: the result is not a generic measure of all cybersecurity work, but a performance figure from a controlled evaluation designed to test multi step cyber operations.
Continue reading
