Thirteen papers covering safety, robustness, and red-team. TruthfulQA, BBQ, and Constitutional AI set the agenda: truthfulness, bias, and an alignment recipe that scales without 100K human labels. GCG, Sclar et al. on prompt sensitivity, HarmBench, and the sycophancy paper showed how easily things break when you push. The 2024 to 2025 results, reward hacking, sandbagging, and alignment faking, are the ones to read before claiming a model is safe.
Two 2026 results change how you run the checks. The emergent misalignment paper (Nature, January 2026) found that finetuning on one narrow task, writing insecure code, produced misaligned answers on unrelated prompts in up to 50% of cases, so rerun the broad safety suite after every finetune, not only the task eval. "A Coin Flip for Safety" audited the LLM judges behind HarmBench-style attack success rates against 6,642 human labels and found accuracy near chance once the attack and victim model change, as low as 45% for the HarmBench classifier on one pairing. Validate the judge on your own attack and victim mix before you trust a robustness number.