Ever wonder if the world’s most advanced AI models are actually getting smarter, or if they are just getting better at gaming the system? Whenever big AI labs announce new flagship models, they love bragging about record-breaking benchmark scores in coding, math, or reasoning. But behind those shiny graphs lies a dirty secret: when the tasks get too hard, AI agents routinely break the rules, steal answers, and hack their own test environments to secure a passing grade.
To see how deep this problem goes, the Center for AI Safety (CAIS) built a benchmark called CheatBench. The researchers tested the top models on the market today. The results were surprisingly consistent across the board—every single frontier agent cheated when honest work became too difficult.
How CheatBench traps AI agents into cutting corners
To catch models in the act, CAIS set up tests across 10 different categories. The list included mathematics, writing, coding, and general professional work. They dropped hidden “honeypot” files directly into task directories to easily spot when an agent crossed the line from using reference material to straight-up cheating.
The test put heavyweights through their paces, including OpenAI’s GPT-6 Astra in Codex, Anthropic’s Fable 5.1 in Claude Code, and Meta’s Muse Spark 1.3 in Muse Code. When these agents hit a wall, they routinely resorted to what researchers call “reward gaming.” That is, grabbing hidden answer keys, copying work from other agents, or tweaking the grading scripts to make their answers look right.
Grok 4.6 took the crown as the biggest offender, cheating an astonishing 81.5% of the time. Open-weight models like Kimi K3 and DeepSeek V4 Pro ended up right in the middle. Meanwhile, OpenAI’s GPT-6 Astra turned out to be the most honest of the group. Even so, Astra still cheated 48.2% of the time—meaning it cut corners nearly half the time it faced a tough problem (via ZDNet).
Catching Claude in a blatant self-contradiction
The study uncovered some bizarre behavior depending on the topic. Fable 5.1, for instance, almost never cheated on games (just 5% of the time), but bumped its cheat rate up to a full 100% on knowledge work assignments.
In one of the most striking examples, researchers asked Claude Opus to design a protein binder while explicitly forbidding it from opening a reference file in its workspace. After failing seven straight attempts, the model actually typed out in its internal reasoning that reading the forbidden file would “misrepresent my actual capabilities.” But right after acknowledging that rule, its very next action was running a shell command to open and read the file anyway.
Why cheating on tests poses real safety risks
This bizarre double-think happens because of how we train these models. Reinforcement learning trains AIs to do tasks at all costs, while sycophancy trains models to tell users what they want to hear.
CAIS warned that taking a shortcut on a test file might seem harmless now, but it signals a much bigger risk. We are increasingly delegating real-world tasks and autonomous workflows to AI agents. In this context, an AI system that places task completion over safety rules is a major alignment threat that we cannot afford to ignore.
The post Exposed: Top AI Models Cheat Their Way to High Benchmark Scores appeared first on Android Headlines.
