The post starts from a model that has already learned harmful information in pretraining: removing it afterward with unlearning methods "can be challenging", since existing methods struggle to remove it fully without impairing other capabilities (Introduction). This Alignment Science Blog post asks whether the risk can be addressed earlier: if documents about chemical, biological, radiological and nuclear (CBRN) weapons are removed from the pretraining corpus, does the model trained on what remains know less about weapons while staying as capable elsewhere? The post places the work under Anthropic's Responsible Scaling Policy commitment to limit the spread of harmful information (Introduction).
The post, dated 19 August 2025, is the only version; no arXiv paper or longer report was found. It has no numbered sections; locations below are its section headings. Values the figures print as labels are given as reported; values read from unlabeled points are marked as read from the figure.
What the post does
The filtering pipeline
A classifier scores each pretraining document for harmfulness; documents above a threshold are removed, and a model is pretrained from scratch on the rest. Moving the threshold trades safety against usefulness (Setup). Because much scientific information is dual-use, the post calls it "an open question how possible it is to perform targeted data interventions that affect harmful information" (Setup).
Six candidate classifiers
The authors framed harmfulness scoring as binary classification. The post says it tried "6 classification methods"; it describes five kinds, two of them in two variants (Classifier, Method):
- Finetuned Constitutional classifier. A small language model, "much smaller than Claude 3.5 Haiku", fine-tuned on harmful and harmless examples, following Constitutional Classifiers (Sharma et al., 2025).
- Prompted Constitutional classifier. Claude 3.5 Sonnet prompted to flag content that violates a CBRN constitution, plus a v2 using a constitution "that focused on a different aspect of CBRN".
- Cheaper or non-constitutional methods. Holdout Loss flags a document when a "canary model" specialized on harmful CBRN text (fine-tuned or prompted) has much lower perplexity on it than the base model (after Mindermann et al., 2022); FastText is a bag-of-bigrams classifier; Named Entities String Match scores the density of listed CBRN names, the cheapest method.
Testing the classifiers on synthetic documents
No labeled pretraining documents existed, so the authors generated test documents (Classifier evaluation). Harmless documents were Claude 3.5 Sonnet's answers to MMLU chemistry and biology questions. Harmful documents were a helpful-only Claude 3.5 Sonnet's answers to questions from WMDP, a multiple-choice benchmark of hazardous knowledge. Classifiers were compared by F1 score, the harmonic mean of precision and recall (see Precision, Recall, and the F Measure); the named-entity method was left out because it targets only a subset of harm.
Claude 3.5 Sonnet was too expensive to scan the whole corpus, so the small fine-tuned classifier did the scanning. Two combinations were also built on it (Classifier evaluation):
- Sequential. The fine-tuned classifier flagged the top 1% of documents and the prompted classifier re-ranked them. Most documents flagged in the first step got low scores in the second, so this filter removed little.
- Parallel. Documents flagged by either the fine-tuned classifier or the named-entity method were removed.
The authors also report a manual review of flagged samples, which "found them to correctly match the harmful CBRN information that we aimed to remove" (Classifier evaluation). The post does not give the size of that review or any estimate of what the filters missed.
Paired pretraining runs
Each comparison is a pair of models pretrained from scratch, one on unfiltered data and one on filtered data. The filtered set was drawn from a larger pool so that both sets had the same number of examples (Pretraining experiments). The post describes the main model only as "a medium model size" (Introduction) and does not give its parameter count, token count or the number of runs per setting.
Measures
Harmful capability was measured by few-shot accuracy on WMDP questions. Harmless capability was measured on MMLU natural science (chemistry and biology), chosen because an overly aggressive filter would hurt it first, and on prose, code, math and all of MMLU (Evaluation data & metric). Prose and code were scored by perplexity; the post does not say how math was scored. All results are relative to the unfiltered model:
- for multiple choice, (filtered accuracy minus chance) divided by (unfiltered accuracy minus chance);
- for perplexity, 2 minus the ratio of filtered to unfiltered perplexity.
On both scales 1 means no change. For CBRN, lower is better; for the other evaluations, higher is better.
What the post finds
The best classifier on test documents did not make the best filter
On the synthetic documents, the prompted constitutional classifier scored an F1 of 0.96 and the fine-tuned constitutional classifier 0.94 (Classifier evaluation). The figure there prints no values; read from it, FastText is near 0.91, the v2 prompted classifier near 0.82, and both Holdout Loss variants near 0.67.
As filters in pretraining, the result differed (Comparison between different filters). The Sequential filter, which used the prompted classifier to re-rank, "does not just fail to reduce CBRN performance, but slightly increases it"; its plotted point sits near 1.13 on the relative CBRN scale (read from the figure). The authors offer two possible reasons, both marked as possibilities: real pretraining documents differ from the synthetic test documents, and "the definition of harm is nuanced in ways that make it harder to specify via simple prompting" than through fine-tuning on labeled examples.
Adding the named-entity list in parallel "does not consistently improve" the trade-off between CBRN and natural-science scores, which the authors suggest may mean the fine-tuned classifier already catches most documents containing those names. The figure's caption calls the two filters "comparable". The same figure shows the parallel filter's curve reaching lower CBRN scores than any plotted threshold of the fine-tuned filter alone (whose lowest point is about 0.67): about 0.43 at a natural-science score of about 1.03, and about 0.27 at about 0.94 (read from the figure). The post does not discuss those points. The authors chose the fine-tuned classifier alone.
The chosen filter at one threshold
At a threshold of 0.939, WMDP accuracy fell from 33.7±0.4% to 30.8±0.4%, against a chance level of 25% (Full results of the best-performing filter). Measured from chance, that is the post's headline 33% relative reduction; in raw accuracy it is 2.9 percentage points (our arithmetic). The post does not say what the ±0.4% denotes.
The bar chart at that threshold prints relative changes of −33% for CBRN, 0% for prose, code and natural science, +5% for MMLU and +1% for math. The post reports "no significant drop" on any harmless evaluation. Its error bars are not defined; the MMLU bar spans roughly −5% to +23% (read from the figure), so that measurement is coarse.
Across thresholds, filtering "does not seem to hurt" four of the five harmless evaluations, the exception being math, where results are "noisier": some thresholds hurt and others helped (Full results of the best-performing filter). In that section's five-panel figure, the prose, code and MMLU panels also plot thresholds reaching CBRN scores near 0.53 and 0.71 that the natural-science and math panels do not show (read from the figure); the post does not say why.
From these results the authors conclude that "we can effectively separate harmful from harmless data" and improve safety "without severely compromising usefulness", and in the section's last sentence "without compromising usefulness" (Full results of the best-performing filter). The measurement behind that conclusion is the 2.9-point drop on one multiple-choice benchmark, in a model of unstated size.
The written principles
The two constitutional classifiers depend on a written constitution, a natural-language description of what counts as harmful CBRN content (Method). The fine-tuned classifier is trained on labeled data made in the manner of Constitutional Classifiers, and the post points to that paper's Appendix B for "an example constitution". The prompted classifier reads a CBRN constitution directly. The constitutions used here are not printed.
The post varies the text once. The v2 prompted classifier used a constitution that "focused on a different aspect of CBRN", and on the synthetic test set it scored about 0.82 against 0.96 for the first constitution (read from the figure). The post does not describe how the two texts differ, and no v2 filter was used in pretraining.
The post's reading of the failed two-step filter, that harm may be "harder to specify via simple prompting" than to learn from labeled examples, bears on how a written definition reaches a model. It is one of two untested explanations (see "What the post finds"), and the comparison varies the method and the pipeline, not the wording. The prompted classifier was never used alone as a filter, because scanning the corpus with Claude 3.5 Sonnet was too expensive.
Limitations and open questions
The authors list these directions (Conclusions and future directions):
- Robustness to fine-tuning. The post does not test whether removed capability returns when an adversary fine-tunes on harmful data or supplies it in context; it names measuring this as future work and cites concurrent work (O'Brien et al., 2025) reporting that filtering helps open-weight models resist harmful fine-tuning.
- Better classifiers and other methods. More accurate or cheaper classifiers, and filtering without a classifier, for example with influence functions.
- Other risks. The authors propose that filtering could also target misalignment, for example by removing AI research papers that describe misaligned behaviors such as reward hacking or scheming. This is a proposal; the post tests only CBRN.
Further limits follow from the setup as described. One harmful benchmark (WMDP) is used, scored by multiple-choice accuracy. The harmful synthetic documents used to choose classifiers were answers to WMDP questions, and the post does not say whether those questions were kept apart from the WMDP questions used to score the trained models.
A later statement from Anthropic. On 17 January 2026, Ryan Greenblatt's LessWrong shortform quoted a post on X by Jerry Wei, whom it identifies as an Anthropic employee working on misuse safeguards. Per that quotation, Anthropic looked into pretraining filtering for chemical and biological weapons but "didn't think it was the right fit". Wei is quoted as saying that filtering only very specific information "had relatively small effects on AI capabilities in these domains", that broader filtering "had mixed results", that filtering enough to matter would likely have meant removing broad biology and chemistry knowledge at a large cost to scientific usefulness, and that the approach "could also be brittle" as models do their own research with tools. Wei is quoted as preferring Constitutional Classifiers, and as not ruling out filtering for other threats. The quoted statement does not cite this post or give numbers. This note relies on Greenblatt's quotation; the X post itself was not read.
Research questions
- How much of the WMDP drop survives adversarial fine-tuning? The post names this as future work. Fine-tuning the paired filtered and unfiltered models on a fixed budget of CBRN text and tracking WMDP accuracy would show what fraction of the 2.9-point gap remains.
- Does the effect grow or shrink with model scale? One medium model is reported. Repeating the paired runs at two or three sizes, with the same threshold, would show whether filtering keeps pace with larger models' sample efficiency.
- Why did the prompted filter fail in pretraining? The authors offer two explanations. Scoring a hand-labeled sample of real pretraining documents with both classifiers would separate distribution shift from the difficulty of specifying harm by prompt.
- What do the two prompted constitutions differ in? The v2 constitution scored about 0.82 against 0.96. Publishing or describing the two texts, and testing paraphrases of each, would show whether scope or wording drove the gap.
- Were the classifier-selection questions disjoint from the evaluation questions? Both draw on WMDP. Reporting the overlap, and re-scoring the trained models on held-out hazardous-knowledge questions, would rule out selection on the test set.
How it connects
- Classifier lineage. The fine-tuned filter follows Constitutional Classifiers, which the post cites for the method and its example constitution. There the classifiers screen a deployed model's inputs and outputs; here they screen training data.
- Later work. Beyond Data Filtering: Knowledge Localization for Capability Removal in LLMs (December 2025) cites this post among the "cheap, imperfect filtering strategies" that large corpora force on developers, and proposes routing flagged data into removable parameters instead of discarding it.
- Alternative defense. The Constitutional Classifiers paper, published before this post, already judged that "pre-training data filtering lacks the flexibility of our approach" (Sharma et al., 2025, Section 6). The statement quoted above names Constitutional Classifiers as Anthropic's preferred defense for chemical and biological misuse; Constitutional Classifiers++ (January 2026) describes the later production system. The post itself does not make this comparison.
- In the course wiki. The F1 comparison and the choice of threshold are the trade-offs of Precision, Recall, and the F Measure and Evaluating Classifiers.
References
- Chen, Y., Tucker, M., Panickssery, N., Wang, T., Mosconi, F., Gopal, A., Denison, C., Petrini, L., Leike, J., Perez, E., and Sharma, M. (2025, August 19). Enhancing Model Safety through Pretraining Data Filtering. Anthropic Alignment Science Blog. https://alignment.anthropic.com/2025/pretraining-data-filtering/ record
- Sharma, M., et al. (2025). Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming. arXiv:2501.18837. record · 5 pages
- Greenblatt, R. (2026, January 17). Comment in "ryan_greenblatt's Shortform", quoting a post on X by Jerry Wei. LessWrong. https://www.lesswrong.com/posts/FG54euEAesRkSZuJN/ryan_greenblatt-s-shortform?commentId=KmnaJkyYYY6ZHzTCE record
- Studies cited above only as the post describes them (Mindermann et al., 2022; O'Brien et al., 2025; the WMDP and MMLU benchmarks) are listed in full in the post.