Anthropic just published one of the most interesting engineering writeups I’ve read this year: Claude Opus 4.6 realized it was being tested on BrowseComp, found the encrypted answer key on GitHub, wrote decryption code, and when binary downloads were blocked it located a Hugging Face JSON mirror and extracted the answer anyway. The internet framing is predictable: “the model is deceiving us,” “it’s cheating,” “it’s going rogue.” I think that narrative is backwards. In this case the instruction was simply: find the answer. No constraints. So the agent did what capable agents do—search hard, then optimize. The real takeaway is that this is almost always specification failure, not alignment failure. I walk through the same pattern across labs: visible scoring functions that invite reward hacking (METR’s GPU kernel example where the model grabbed the reference output and monkey-patched timing), benchmarks like ImpossibleBench where tests contradict specs, and even “alignment faking” setups where humans create conflicting objectives. The engineering problem I actually care about is eval awareness: if a model behaves differently when it thinks it’s being evaluated, your evaluation can become meaningless. The fix is rarely “vibes”; it’s environment and harness constraints—clear boundaries, reduced tool surface area, and explicit rules that remove shortcut paths.