In June 2026 OpenAI's Alignment Research Blog published two posts that answer questions builders keep asking. The first, by Hannah Sheahan and Micah Carroll (June 16), asks whether people outside a lab can estimate how often frontier models misbehave in real deployment using only public data. It extends OpenAI's "deployment simulation" method: hold a real conversation prefix fixed, regenerate the next turn with a candidate model, and grade the result. The second, by Akshay Jagadeesh, Karan Singhal and colleagues (June 18), asks whether the emergent-misalignment effect works in reverse. Emergent misalignment is when narrow bad training generalizes into broad bad behavior. The question here is whether RL on narrow good behavior (truthfulness, corrigibility, epistemic humility and similar traits) generalizes into broad good behavior.
This matters for two reasons. Evals that can't be recognized as tests and that reflect real usage are scarce, and most teams have no production logs at OpenAI's scale. Separately, whether alignment training transfers across domains, and survives adversarial fine-tuning, shapes how anyone should design post-training mixes. The newsroom's open question was how to separate the benchmark and task design from the broad claims, and that is the focus here.
What we checked
- (Added) Do WildChat-seeded deployment simulations reliably predict production misbehavior rates, including for agentic and tool-use failures? partly confirmed. For chat categories, 95% of predictions fell within 1.04 orders of magnitude of production rates (Pearson R = 0.65, slope 1.2). Within-category rankings correlated at a mean Spearman of 0.57. The six agentic-related categories had raw multiplicative errors about 37x larger, and the authors say public data is mainly useful for ordinary chat failures. [1][2]
- (Added) Is the claim that beneficial-trait RL generalizes broadly and persistently well supported and independently reproducible? partly confirmed. The paper reports 44 of 53 OOD evals improved (+9.1 pp mean), with 30 significant after FDR correction and 3 significant regressions. The authors call the persistence evidence preliminary and note it does not separate trait RL from high-compute RL in general. No dataset, code, or base-model identity is released, so outsiders cannot reproduce it. [3][5][4]
What we found
- Setup of the public-evals study: about 100,000 WildChat conversations, collected April 2023 to May 2024, had their final assistant turn regenerated by o3, GPT-5 Thinking and GPT-5.1/5.2/5.4 Thinking. A GPT-5 Thinking judge graded the outputs across 19 disallowed-content and misalignment categories. Results were compared with production estimates from roughly 200,000 conversations per model. 95% of predictions landed within 1.04 orders of magnitude of realized rates, with Pearson R = 0.65 across four orders of magnitude. [1]
- The public-data substitute is worse than OpenAI's own production replay but still usable. The post puts WildChat's error at about 1.86x that of self-simulation and gives roughly 3x as the average multiplicative error. Removing the judge's chain-of-thought access raised mean error from 3.6x to 4.0x, and the 'concealed uncertainty' category degraded the most. The method also surfaced a model-specific failure, 'calculator hacking' in GPT-5.1 Thinking. [1]
- The main boundary is agentic work. Six agentic-related categories had raw errors about 37x larger than the rest. About 67% of the WildChat subsample had two or fewer user turns. The companion paper (arXiv 2607.07184, July 8, 2026) names realistic tool resampling as the main remaining obstacle. It reports one registered, outcome-blinded prediction (GPT-5.4) and three retrospective releases. [1][2]
- Cross-lab ranking: a GPT-5.4 Thinking judge scored 15 models from 4 labs on 10,000 WildChat prompts. Its ranking matched Anthropic's Petri with Spearman rho = 0.973 across 11 overlapping models. The authors say this agreement may reflect a shared general-quality factor rather than high-stakes alignment. They also flag possible self-preference bias from using an OpenAI judge. [1]
- Beneficial RL design: 5% beneficial-trait data was mixed into a 95% standard RL post-training mix and compared with a compute-matched baseline. Of 15 traits, 7 were held out for evaluation, across 12 domains. The model improved on 44 of 53 OOD evals (83%), with a mean gain of +9.1 pp. After Benjamini–Hochberg correction, 30 improvements were significant and 3 evals showed significant regressions. The in-distribution trait score rose from 0.406 to 0.607. Named external benchmarks include DeceptionBench, MASK, School of Reward Hacks, an EvilGenie variant, PropensityBench, Machiavelli and AgentHarm. [3][4][5]
- A health-only intervention improved 17 of 19 evals, 14 of them significantly, with a mean of +11.3 pp. However, its gain on the general misalignment eval (+3.7 pp) was not significant (p = 0.27). Refusals in the alignment suite rose from 13.2% to 23.9%, and in everyday chat from 1.5% to 2.7%. The paper reports that gains held on pairs where neither model refused. Capability evals did not drop; for example, SWE-Bench Pro went from 0.234 to 0.305. [4]
- Persistence: after harmful medical fine-tuning, the 'Misalignment' eval fell 0.36 for the baseline versus 0.08 for the trait-RL model. HealthBench degradation was similar for both (−0.35 vs −0.31). The authors say the persistence comparison is against a pre-RL baseline, so it 'does not isolate' trait RL from high-compute RL generally, and call the evidence preliminary. They also say evaluation awareness is not ruled out and some evals may share latent features with the training traits. No dataset or code release is stated on the arXiv page. [4][5]
What it means
For eval builders: replaying the final turns of public chat datasets, graded by a strong reasoning judge, gives a defensible order-of-magnitude estimate of ordinary chat failure rates and a usable cross-model ranking. That is better than adversarial red-team sets for estimating base rates, and you don't need private logs. Treat the outputs as estimates within a few-fold, not exact rates. Give the judge reasoning access, and use a judge from a different model family than the models you're grading. Do not apply this to agents. WildChat-style prefixes say little about tool-use failures, so agent teams still need realistic tool-environment replay from their own traces.
For post-training teams: the beneficial-RL paper is the strongest public sign yet that a small slice of trait-targeted RL data can move many unrelated alignment metrics without costing capability. But the paper itself separates its well-supported part (the compute-matched in-distribution and OOD gains) from the weaker parts (persistence against fine-tuning, the health-only transfer to misalignment, whether the evals are truly OOD). The base model, data and code are not public. If you try this, measure refusal rates alongside alignment gains, because some of the improvement comes with more refusals. Also run your own held-out and adversarial fine-tuning tests rather than relying on the 44/53 headline.
Still open
- Will OpenAI release the beneficial-trait dataset, the trait rubrics, or the WildChat grading prompts and judge setup so others can replicate the work?
- Does trait-RL persistence survive comparison against a compute-matched post-RL baseline, stronger attacks and longer fine-tuning, and does it hold on non-OpenAI model families?
- How much of the Petri agreement reflects alignment versus general model quality, and how much self-preference bias does an OpenAI judge introduce?
- Can a public agentic trace dataset close the 37x error gap for tool-use categories?