GPT-6 Astra posts near-perfect reverse-engineering scores, a capability that now needs governance more than applause

We only have a short video note on this one, so we will keep it short. The note says OpenAI's GPT-6 Astra reached an 88% success rate on reverse-engineering tasks in a single attempt, and 99.2% within four attempts, on a benchmark it calls 'S Bench'. Reverse engineering means taking apart compiled software to work out how it functions without access to its source code.

What we cannot verify matters as much as what we can. The note gives no detail on the benchmark's design, task difficulty, how the model was run, or who ran the evaluation. We are not adding to it. The gap between 88% on the first try and 99.2% within four tries suggests a test close to saturation. As we have argued about cybersecurity benchmarks before, saturated scores say little about real-world capability. What separates models is expert-level work done without guidance, under independent evaluation.

Our reading: if the figures hold up, this is a real step in a dual-use skill. Reverse engineering underpins malware analysis, vulnerability research and interoperability work. The same skill also helps attackers find and exploit flaws. The right frame is capability and risk to be governed, not a leaderboard win. We want to see independent results on harder tests before treating 'near-perfect' as more than a headline.