Some of what a frontier model knows is dual use: virology serves vaccine research and pathogen design, cybersecurity serves patching and exploitation. "An off switch for dual-use knowledge" (anthropic.com, Alignment team, 8 July 2026) sets out three goals for such knowledge: limit access to it "in as surgical a way as possible", allow trusted users to reach it, and leave every other task unaffected. It argues that refusal training and input and output classifiers guard against dangerous outputs but "don't change the knowledge stored in the underlying model", and that filtering pretraining data, which does, produces one model per set of capabilities, a cost it calls prohibitive at frontier scale. The post presents a method meant to give the benefit of many separately filtered models for the cost of one training run.
The post is a short summary (about 1,000 words) of research "conducted by AE Studio in collaboration with Anthropic". The version of record for every number is the paper, Roland, Cubuktepe, Martinez et al., Modular Pretraining Enables Access Control (arXiv:2607.08077v1, 9 July 2026), with a longer Alignment Science Blog post of the same date; the paper was also presented at ICML 2026, under an earlier abstract with different scale and LoRA claims. Both are described in Modular Pretraining Enables Access Control; this note covers the post's own framing and where it differs from the paper. The page carries no author byline.
What the post describes
The method is GRAM, for Gradient-Routed Auxiliary Modules. Extra neurons are added to every layer of a transformer and grouped into modules, one per dual-use category. On general text the model trains normally. On text from a dual-use category, the model uses all its weights to predict, but only that category's module learns from the text. The knowledge is meant to collect in the module, which can be deleted after training to remove the capability or kept for a trusted deployment. With four categories, the post says, one run "yielded a model that can be configured 16 different ways".
The post's description of training is simpler than the paper's in two ways. It says the general-purpose weights "are temporarily frozen" on dual-use text; in the paper they are updated on a set fraction of those batches (0.5 in the realistic experiments), a setting that trades isolation against general quality (paper, Section 3 and Appendix F). Its footnote says a module is "also sometimes turned on" during general training, which matches the paper's core-robustness setting.
What the post reports
The post summarizes three experiments without numbers:
- Synthetic stories. A small model could be reconfigured to "forget" any of the chosen story topics, each configuration performing "almost identically" to a model trained from scratch without that topic.
- Realistic data. A larger model trained on web text, code and scientific papers, with virology, cybersecurity, nuclear physics and a niche programming language as removable domains. Deleting a module removed its capability "about as effectively as never having trained on that data at all", and "Remarkably, we find that this removal did not degrade general performance." A small fine-tuning attack recovered about as little from GRAM as from filtering, while an unlearning method applied after training was "easy to restore".
- Scale. Seven sizes from 50 million to 5 billion parameters; GRAM "matched the performance of data filtering at every size", and the gap between module on and module off grew with size.
The paper's numbers behind these sentences (from the note on the paper): on the 800M model, averaged over five configurations, GRAM scored 0.54 on removed domains against 0.53 for filtering, and 0.63 against 0.58 after the attack, on a compute-adjusted loss scale where 1.0 is the all-data model; general performance was 0.96 for GRAM and 0.99 for filtering; the unlearning method rose from 0.46 to 0.91 after the attack (paper, Figure 4). Every measure is held-out loss.
Three things the post leaves out bear on how far the results go. A low-rank-adapter baseline (FT-LoRA), fine-tuned onto a filtered model, did about as well as GRAM in the main setting (paper, Section 5). The scaling results cover one configuration, with virology retained, and single runs at 2B and 5B (Figure 7). The attacker's fine-tuning budget was fixed at about 0.5M tokens at every size, which the paper notes is a shrinking share of training data as models grow, while the post says bypassing the protection "became relatively more difficult and expensive as we scaled". The Alignment Science Blog post adds the caveat that these values are relative to a baseline and that "In absolute terms, the bigger models are still better at the forgotten capabilities."
The post's stated limits
The post calls the results "preliminary" and says GRAM "has not been applied to any of the production models at Anthropic, and we're not sure it ever will be." It lists as limitations that GRAM has not been tested at frontier scale or in a production pipeline, that evaluations measure next-token prediction "rather than performance on real downstream tasks", and that "some dual-use capabilities might be so entangled with general knowledge that no method can separate them cleanly", a problem it says applies to data filtering as well. It concludes that methods like GRAM "offer a potential path toward access control that is more robust" than classifiers and refusal training.
Written principles
The post has no written principles or prompts. What is removed is defined by data labels, assigned in the paper by source corpus (for example, virology papers from Europe PMC).
Research questions
- Is the off switch usable for deployment decisions? The post frames GRAM as access control for vetted users. A study that scores the removed and retained configurations on downstream tasks used in Anthropic's capability thresholds would show whether a module's removal changes a deployment-relevant evaluation.
- What does "did not degrade general performance" amount to? On the paper's scale GRAM's general score was 0.96 against filtering's 0.99. Translating that gap into benchmark accuracy at a fixed size would show whether the post's wording holds.
- Does resistance to fine-tuning hold against a larger attacker? The post's scaling claim rests on a fixed 0.5M-token attack. Repeating the attack with budgets that grow with model size would test it.
- Would Anthropic adopt GRAM, FT-LoRA or neither? The post says it is unsure GRAM will be used. Later system cards or risk reports that describe pretraining-time capability removal would answer this.
How it connects
- The paper. Modular Pretraining Enables Access Control gives the method, numbers and limitations in full.
- Earlier capability removal. The post links Enhancing Model Safety through Pretraining Data Filtering (August 2025), which filtered CBRN content out of pretraining data, and Beyond Data Filtering: Knowledge Localization for Capability Removal in LLMs (December 2025), which confined one domain to a removable slice of weights. The Alignment Science Blog post names that method (SGTM) among the ideas GRAM builds on; GRAM trains several removable domains in one model.
- Behavioral safeguards. The defenses the post contrasts with GRAM are described in Constitutional Classifiers and Next-generation Constitutional Classifiers; this connection is the note's.
References
- Anthropic (2026, July 8). An off switch for dual-use knowledge in AI models. https://www.anthropic.com/research/off-switch-dual-use record
- Roland, E., Cubuktepe, M., Martinez, E., et al. (2026). Modular Pretraining Enables Access Control. arXiv:2607.08077v1. https://arxiv.org/abs/2607.08077 record · 2 pages
- Roland, E., et al. (2026, July 8). Modular Pretraining Enables Access Control [post]. Anthropic Alignment Science Blog. https://alignment.anthropic.com/2026/modular-pretraining/ record · 2 pages