Safety work moves into testing and model internals
By Nathan Benaich (Air Street Capital) · 2024 report
The 2024 report documented practical work on evaluating and controlling models alongside a changing public debate about AI risk. New defenses addressed specific failure modes, and interpretability researchers identified features inside a frontier model. Neither established that complex systems had become fully understood or reliably safe.
Defenses target how models handle malicious requests
The report covered instruction hierarchies, cautionary warnings, and circuit breakers that attempted to interrupt harmful representations. New red-team benchmarks made defenses easier to compare. It also retained the central limitation: success against a collection of known attacks does not guarantee resistance to a new one.
Anthropic used sparse autoencoders to separate Claude 3 Sonnet’s internal activations into more interpretable components. Adjusting a feature could influence the model’s output, as the widely discussed Golden Gate Bridge example showed. This offered a way to study parts of a model’s behavior without amounting to a complete explanation of the model.
The UK AI Safety Institute released Inspect to support evaluations of knowledge, reasoning, and autonomous capabilities. The report described growing international coordination, while questioning how much evaluation work would depend on voluntary cooperation from model developers. Institutional capacity and reliable model access remained separate issues.
Sparse autoencoders offered a way to separate internal activations into components that researchers could interpret and manipulate. The Golden Gate Bridge example showed that such a feature could influence outputs. It did not mean the model’s entire computation, reliability, or safety properties had become transparent.
Evidence you can use
AI safety in the 2024 report
Historical snapshot: October 2024. Dates and populations are specified per row.
These rows identify methods, models, and institutions rather than comparable safety scores. The report did not claim that jailbreak defenses were universal or that interpreting selected features explained all model behavior.
Sparse autoencoders identified interpretable components in Claude 3 Sonnet’s activations, and adjusting features could influence outputs. This was progress in studying model internals, not a complete account of how the model worked.
The report described new defenses and evaluations, but also debate about whether benchmark success would transfer to strong new attacks. It did not treat jailbreak resistance as solved.
It distinguished higher-priority instructions from lower-priority or untrusted text. The report described this as one defense against treating every piece of text as equally authoritative.
They attempted to interrupt harmful internal representations or behavior. The report treated them as a defense approach to evaluate, rather than proof that all future attacks would fail.
A defense tested on a selected attack set could still face new strategies or inputs. The report preserved the distinction between measured benchmark resistance and a general guarantee.
It showed that adjusting an interpretable component could influence Claude’s output. That linked an internal feature to behavior without providing a complete explanation of every model decision.
No. The framework supported evaluation work, while model access could still depend on developer cooperation. Evaluation tools and the authority or opportunity to use them were separate matters.
Historical snapshot published October 10, 2024. This web edition was prepared on October 10, 2026 from the online deck and original launch posts. Findings and forecasts retain their original time frame.
2024 report, slide 178 (PDF page 179)Original 2024 report. Printed slide numbers are one lower than PDF page numbers because the cover is unnumbered.
2024 report, slide 181 (PDF page 182)Original 2024 report. Printed slide numbers are one lower than PDF page numbers because the cover is unnumbered.
2024 report, slide 197 (PDF page 198)Original 2024 report. Printed slide numbers are one lower than PDF page numbers because the cover is unnumbered.
Benaich, Nathan. “Safety work moves into testing and model internals.” State of AI Report 2024. Historical report snapshot; web edition prepared 2026-10-10.