Enigma (Product): news timeline & CVEs

Where Do the Tokens Go? Understanding and Reducing Costs in LLM Agents for Vulnerability Discovery

A 200-trace CyberGym study finds reasoning dominates token cost, and AVRI lowers it.

Researchers open-coded 200 CyberGym traces from Codex, OpenCode, Cybench, and EnIGMA under an unaided baseline and four existing efficiency methods. Agents vary widely in success and cost, and higher spending does not consistently improve outcomes. Code localization, vulnerability reasoning, and trigger design account for 60.4% of tokens, while only 24.4% of matched comparisons preserve success at lower cost. Their AVRI interface, using a persistent Bidirectional Evidence Trace, reduces cost 18.0% for Codex and 23.7% for OpenCode on 20 tasks without lowering success.

arXiv cs.CROct 8, 09:47 UTC · 8d agoAI research

Astra and Opus just passed Turing’s other test

Cryptanalysts used OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5 to crack two unsolved WWII Enigma messages.

Cryptanalysts report that OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5 decoded two archival Enigma messages that had remained unsolved. Developer Carter Leffen prompted Astra to find and break a message unsolved since 2005; Frode Weirerud later validated the plaintext. On September 21, Jack Willis said Claude Opus 5 broke a different message after he supplied an officer's known signature as a crib. Weirerud says seven Enigma messages are still unbroken, plus one whose plaintext is known but whose settings are not.

TechCrunch · AISep 25, 17:24 UTC · 21d agoAI research

Enigma

Timeline2 articles

Appears with

Where Do the Tokens Go? Understanding and Reducing Costs in LLM Agents for Vulnerability Discovery

Astra and Opus just passed Turing’s other test