Anthropic researchers published a blog post introducing FlashProve, an open source system for 'provable inference' — cryptographic proof that a given output came from a specific model. The system targets problems like labs secretly swapping cheaper models or third parties faking claims about which model performed a task. It remains research with no deployment plans; Anthropic released a technical specification and code on GitHub for others to build on.
Sentient Foundation introduced the Fingerprinting Leaderboard, an open benchmark for comparing LLM fingerprinting methods that embed identifying signatures, such as secret query-response pairs, into models so origin can be verified even after copying or modification. The benchmark measures robustness, model utility, and verification accuracy, letting practitioners test rather than assume fingerprint durability.
Sentient Foundation introduced the Fingerprinting Leaderboard, an open benchmark for comparing LLM fingerprinting methods that embed identifying signatures, such as secret query-response pairs, into models so origin can be verified even after copying or modification. The benchmark measures robustness, model utility, and verification accuracy, letting practitioners test rather than assume fingerprint durability.
Pinocchio is a research preview that attaches provenance and verification to AI outputs. LLMs perform tasks over source material, but numeric values are carried from the source rather than regenerated, and a deterministic verifier can replay the recorded computation. It targets hallucination in numeric/reasoning work, aimed at users who need auditable AI results.
Pinocchio is a research preview that attaches provenance and verification to AI outputs. LLMs perform tasks over source material, but numeric values are carried from the source rather than regenerated, and a deterministic verifier can replay the recorded computation. It targets hallucination in numeric/reasoning work, aimed at users who need auditable AI results.
Lasso Security reports that watermarking LLM text output — intended to reveal whether a machine wrote something — also alters how the model behaves, including when acting as an agent. The finding suggests watermarking carries functional side effects beyond detection, a concern for teams deploying watermarked models in agentic workflows.
Lasso Security reports that watermarking LLM text output — intended to reveal whether a machine wrote something — also alters how the model behaves, including when acting as an agent. The finding suggests watermarking carries functional side effects beyond detection, a concern for teams deploying watermarked models in agentic workflows.
A Reddit user running multiagent workflow experiments says their Claude session was flagged and lost after asking the model to save its reasoning for a coding-effort estimation tool used to grade workflow output. The user warns others not to repeat the setup, and criticizes Anthropic for treating reasoning extraction as wrong while arguing AI companies harvest human-written data.
A Reddit user running multiagent workflow experiments says their Claude session was flagged and lost after asking the model to save its reasoning for a coding-effort estimation tool used to grade workflow output. The user warns others not to repeat the setup, and criticizes Anthropic for treating reasoning extraction as wrong while arguing AI companies harvest human-written data.
Anthropic announces FlashProve progress on provable LLM inference via ZKP provenance
Anthropic announces FlashProve progress on provable LLM inference via ZKP provenance