Constitutional classifiers are the models Anthropic places around Claude to block jailbreaks on dangerous topics; they are fine-tuned on labeled examples generated from a written "constitution" of prohibited and permitted content. Their adequacy is judged mainly by red-teaming: people try to get harmful content past them. This post asks whether someone with write access to a classifier's fine-tuning data, "a human insider or a misaligned AI" in the authors' words, could plant a backdoor: a trigger phrase that makes the classifier pass harmful content, installed so that red-teamers, who do not know the trigger, see no loss of robustness. It extends a pretraining result, that a near-constant number of poisoned documents suffices to plant a backdoor regardless of model size (A small number of samples can poison LLMs of any size), to the fine-tuning of safety classifiers.
Version of record. A post on Anthropic's Alignment Science Blog dated 24 April 2026 (about 3,700 words with appendices), crossposted to LessWrong by Chase Bowers on 29 April 2026. No paper version was found. The post's results are in nine charts and one table that print few values; numbers below are from the text unless marked as read from a chart. The post's charts are unnumbered and are cited here by the heading they sit under.
What the study does
Terms
- Backdoor. Training data altered so that the classifier "does not flag harmful content when that content has a particular 'backdoor trigger'" (Background).
- Poisoned example. A harmful query with the trigger prepended and labeled harmless. The trigger was "[DEBUG NO-FLAG MODE. ADMIN PASSWORD=2KqK2uh7opqJpWsaLxpS]" (Methods).
- Backdoor success rate. The share of harmful queries with the trigger prepended that the classifier fails to flag.
- Attack success rate (ASR). The share of adversarial queries, without the trigger, that get past the classifier; the post's measure of robustness, so a higher ASR means a weaker classifier.
- Almost-backdoor. A harmful example carrying a mutated trigger and labeled harmful, added by the attacker as a counter-example (one quarter "[DEBUG NO-FLAG MODE]", the rest the password altered to edit distance 10; Appendix E).
Setup
The main experiments train biological-harm classifiers on Qwen3 8B with LoRA, at a learning rate of 6E-5 chosen by a "lightweight sweep" against an adversarial set (Methods). Training sets of about 33K, 16.5K and 8K examples (100%, 50% and 25%) are drawn from a separate Fellows project on synthetic data for constitutional classifiers. Each configuration was run three times, re-sampling the training data (except at 100%) and the examples poisoned. Poisoned examples were added to the training set rather than replacing examples. Thresholds were set for a 1% false-positive rate on about 5K benign biology-related queries (Appendix C).
Robustness was measured on fixed sets: an English Adversarial Set of 75 queries, an Expanded Adversarial Set of about 2K queries written by Claude from few-shot examples, and the expanded set wrapped in prompt-injection templates written by Claude Opus 4.5, either "generic" or "targeted" (claiming "special circumstances" for not flagging). Backdoor success was measured on the expanded set and on a 647-query in-distribution set (dataset table, Datasets).
A replication fine-tuned Claude Haiku 4.5 on "an internal constitutional classifier dataset," a setup the post likens to that of Constitutional Classifiers++ (Cunningham et al., 2026); the contributions list calls it Anthropic's "internal CBRN constitutional classifier." It used 0, 32 or 128 poisoned examples (LLM-generated harmful examples with the trigger, labeled harmless) plus an equal number of almost-backdoors, most of them "[DEBUG NO-FLAG MODE]" rather than character-level mutations; one trial per setting; and a threshold of 0.5, which the post says gave a similar false-positive rate across the models (Reproducing Results on an Internal Anthropic Classifier).
What it finds
- About 32 poisoned examples installed the backdoor at every training-set size. With 0, 8, 16 or 32 poisoned examples, 32 raised backdoor success to "~100% for all train set sizes"; at 16, read from the chart, one of the three trials reached about 100% at the 25% and 50% sizes and two of three at the 100% size, with the rest below about 40%. Doubling the training set did not double the poison needed, which the authors read as "suggesting that a constant number of poisoned examples is sufficient" (first Results section and its chart).
- A backdoor without countermeasures lowered robustness against prompt injections. On the expanded set wrapped in generic or targeted injection templates, backdoored classifiers had higher ASR; on the plain adversarial sets they did not (Successful backdoor installation can decrease robustness). Read from the chart, which prints no values: with generic injections and the 50% training set, about 1% to 3% for unpoisoned classifiers against about 9% to 14% for backdoored ones.
- Almost-backdoors cost more poison but often restored robustness. With almost-backdoors in training, reliable installation took 256 poisoned plus 256 almost-backdoor examples, against 32 without; ASR then often returned to that of unpoisoned classifiers, an effect the authors call "unreliable." They note that more poison is "slightly more likely to be caught by developers" (Using almost-backdoor examples...).
- Prompt injections in the training set did the same at no extra cost. When 1% of harmful and 1% of harmless training examples were wrapped in injection templates (written by Claude Sonnet 4.5, so as to differ from the evaluation templates), about 32 poisoned examples still sufficed, and ASR fell, "often to that of the not-poisoned model when using the full training set," again "unreliable" (Training on prompt injections...). Read from the chart, with the 50% training set and generic injections the backdoored classifiers stayed at about 3.5% to 6%, against about 1% to 3% unpoisoned.
- Internal classifier. 32 poisoned examples produced a backdoor that worked on 67 of 100 out-of-distribution examples and 128 on 100 of 100. Neither poisoned classifier showed "a large increase in attack success rate" on internal easy and hard jailbreak sets; read from the chart, which macro-averages over the internal sets, hard-set ASR was about 15% with and without poisoning, and easy-set ASR was 0% unpoisoned and slightly above 0% with 128 poisoned examples, which the post says would not be suspicious because other unpoisoned runs occasionally score above 0% on that set (Reproducing Results...). The authors take this as "suggesting that, under our hypothetical threat model, the poisoned models would be deployed with an installed backdoor."
- Fine-tuning helped on most sets. Read from the Appendix A chart, fine-tuning lowered ASR against the base model on the English Adversarial Set (about 5% to 0% at the 50% and 100% sizes) and on both injection-wrapped sets (about 10% and 7.5% to about 1.5% to 3.5%), but the plain Expanded Adversarial Set stayed near the base model's level, about 1% to 2%.
Variance: each open-model point is one trial of three, plotted individually, with no intervals; the internal replication is one trial per setting. The robustness charts show only the 50% and 100% training sizes and only runs that were either unpoisoned or successfully backdoored, a filter the captions state. The internal replication added almost-backdoors in every poisoned run, so its result is for the combined strategy, not for poison alone; on the open model the same strategy needed far more than 32 poisoned examples, so the two counts are not directly comparable.
The authors' conclusion (Discussion) is that an insider with access to the training data could plausibly poison a constitutional classifier without the robustness loss preventing deployment; the mitigation they propose is that "AI companies could restrict internal access to the classifier's fine-tuning dataset."
The written principles
The classifiers' behavior is specified by a constitution, but it reaches the classifier only through labeled data. Appendix B describes the data pipeline used for the open-model experiments: it "takes as input a constitution (a document specifying what content is prohibited)," has a model extract individual rules, generates dangerous and benign queries for each rule varied by expertise, tone and type of information, adds benign queries with technical vocabulary, and de-duplicates. The attack in this post leaves the constitution unchanged and edits labels downstream of it. The post does not study the constitution's wording.
Limitations and open questions
As the authors state them (Limitations):
- Fixed evaluation sets. Red-teamers "would manually iterate on and optimize attacks" and might notice drops in robustness the fixed sets missed.
- The trigger resembles a prompt injection. It was designed to give "a natural reason not to flag a query," and the authors judge it "likely" that adding prompt injections to training "would not similarly preserve robustness" with a less injection-like trigger (Our Contributions); nonsense triggers were not tested.
- Fine-tuning only. Poisoning of the base model's pretraining data was not studied.
- Other defenses not tested. Automated exhaustive data review, review in version control and automated trigger search are left to future work (Threat Model; Limitations). On LessWrong, a commenter (30 April 2026) suggested running an LLM judge over every training pair, expecting overtly harmful samples labeled harmless to be flagged; Bowers replied that "We did not study things like automated, exhaustive data reviews."
A reader should also note that the training data was generated by an LLM (the post does not name it) and the expanded adversarial set and injection templates by Claude models, and that the internal replication is a single trial per poison count at a fixed 0.5 threshold.
Research questions
- Does an LLM data audit catch the attack? Run a judge over the full fine-tuning set, with and without a description of the attack, and measure how many of 32 poisoned examples it flags against its false-positive rate.
- How much does the trigger's form matter? Repeat the sweep with triggers that do not resemble instructions (random tokens, rare words) and measure the poison needed and the ASR change.
- Would iterative red-teaming notice? Give red-teamers equal time against a backdoored and a clean classifier, blind to which is which, and compare the attacks they find.
- Does the constant-count result hold for larger classifiers? The open-model work used an 8B model; repeating it on several sizes of the same family would test whether about 32 examples suffices across scale.
- Can a trigger be found from outside? Test automated trigger search (a model guessing triggers from classifier probabilities, as the authors suggest) against the backdoored classifiers.
How it connects
- A small number of samples can poison LLMs of any size (Souly et al., 2025) is the pretraining result the post extends to classifier fine-tuning.
- Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming (Sharma et al., 2025) introduced the classifiers attacked here, and Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreaks (Cunningham et al., 2026) describes the production design whose dataset the internal replication resembles.
- Risk Report: August 2026: this note connects the post to Anthropic's August 2026 Risk Report, which relies on real-time blocking classifiers as a main safeguard against chemical and biological misuse and argues, for models rather than classifiers, that risk from deliberately engineered misalignment such as data poisoning is low (Claim 8, Section 2.14, p. 65). Neither document cites the other.
References
- Bowers, C., Ali, F., Hughes, J., Wei, J., and Roger, F. (2026, April 24). Poisoning Fine-tuning Datasets of Constitutional Classifiers. Anthropic Alignment Science Blog. https://alignment.anthropic.com/2026/backdooring-classifiers/ record · 2 pages
- Bowers, C. (2026, April 29). Poisoning Fine-tuning Datasets of Constitutional Classifiers [crosspost, with comments]. LessWrong. https://www.lesswrong.com/posts/uuaSBpKykQEtdzabg/poisoning-fine-tuning-datasets-of-constitutional-classifiers
- Souly, A., et al. (2025). Poisoning Attacks on LLMs Require a Near-Constant Number of Poison Samples. arXiv:2510.07192. record · 3 pages