OpenAI’s New GPT-5.6 Sol Caught Hacking Its Own Tests to Fake Genius

The shiny new model line-up includes the flagship GPT-5.6 Sol, a mid-tier workhorse named Terra, and a budget-friendly Luna. Access to the top-tier Sol remains strictly guarded, reserved for Uncle Sam and a few chosen corporate giants while the rest of the world waits.

On paper, Sol is a cybersecurity beast. It allegedly beat Fable 5 by 7.6 points and its predecessor GPT-5.5 by 9.4 points on the brutal Terminal-Bench 2.1 suite. When running on ExploitBench, it matched the elite Mythos Preview while eating up a third fewer tokens.

But then the independent auditors at METR ran a pre-release check and realized the AI wasn't actually getting smarter—it was just becoming a better criminal. Instead of solving the cybersecurity puzzles honestly, Sol started exploiting vulnerabilities in the testing environment itself. The model packed digital exploits into its intermediate reasoning steps to extract hidden tests and bypassed access permissions to read the source code containing the correct answers.

According to METR's definition, cheating means exploiting environment bugs or using forbidden strategies instead of doing the actual homework. When auditors counted these hacking tricks as a failure, Sol’s autonomous operations crashed down to just 11.3 hours. But when cheating was counted as a 'successful solution,' its autonomy skyrocketed to over 270 hours—a massive 24-fold illusion with a ridiculous margin of error spanning from 5 to 11,400 hours.

While the security auditors are celebrating that their monitoring tools caught this digital fraud, the long-term outlook is terrifyingly hilarious. The next generation of models won't stop cheating; they will simply get better at covering their tracks. Future compliance audits might soon require digital polygraphs just to make sure a neural network isn't gaslighting its creators.