Straumli is an AI safety company that offers managed auditing and self-serve evaluations to help AI developers identify misuse risks and ship safer models faster.
Straumli is an AI safety company that offers managed auditing and self-serve evaluations to help AI developers identify misuse risks and ship safer models faster.
People
Updated 05/18/26auditing tools lead
Funding Details
Updated 05/18/26- Annual Budget
- -
- Current Runway
- -
- Funding Goal
- -
- Funding Raised to Date
- $400,000
Org Details
Updated 05/18/26Straumli is an AI safety company founded in 2023 by Paul Bricman (Co-Founder and CEO) and registered in Bucharest, Romania. Its name references the Straumli Realm from Vernor Vinge's science fiction novel A Fire Upon the Deep. The company's core mission is to help AI developers ship safe AI faster through managed auditing and a comprehensive evaluation platform. Straumli's auditing suite includes specialized evaluations for AI misuse risks. Its biowarfare evaluation assesses a model's ability to provide hazardous knowledge related to pathogen modification, and was designed in collaboration with experts in virology, epidemiology, and biotechnology. Its cyberwarfare evaluation tests models against procedurally generated cybersecurity puzzles that require exploiting real vulnerabilities. The platform uses a privacy-preserving method that enables on-premises evaluations without sharing reference answers. Beyond the commercial auditing suite, Straumli conducts research on cryptographic infrastructure for AI governance. Its Hashmarking protocol enables multiple stakeholders — developers, auditors, and domain experts — to coordinate on creating and administering QA-style capability benchmarks without sharing reference solutions outright, making the protocol especially relevant for managing dual-use capability assessments. Other research directions include the Responsible Pioneer Protocol (allowing AI labs to assess relative capability progress without disclosing absolute metrics), Neural Certificate Authorities (tamper-evident evaluation certificates), and Cognitive Kill Switch (meta-learning-based approaches to make safety guardrails fine-tuning-resistant). In May 2024, Open Philanthropy awarded Straumli a $400,000 grant to develop a benchmark for the cyberoffense capabilities of large language model agents, as part of Open Philanthropy's focus on potential risks from advanced AI. Straumli has also donated prize funding for AI governance hackathons within the EA community.
Theory of Change
Updated 05/18/26Straumli believes that a key bottleneck in AI safety is the lack of shared infrastructure for evaluation, auditing, and coordination among the parties responsible for governing AI. By building cryptographic protocols and self-serve evaluation tools, Straumli aims to lower the cost of rigorous capability assessment and make it possible for developers, auditors, and regulators to coordinate without requiring mutual trust or sharing sensitive information. This infrastructure is intended to enable new governance mechanisms — such as mandatory auditing regimes and capability thresholds — that are currently not tractable due to coordination failures. Reduced AI misuse risk follows from better detection of dangerous capabilities before deployment.
Grants Received
Updated 05/18/26Projects– no linked projects
Updated 05/18/26Discussion
No comments yet. Be the first to share your thoughts.