Auditbench
Auditbench2About
What does Auditbench do?
AuditBench is a benchmark designed to evaluate the effectiveness of AI alignment auditing techniques. It comprises 56 language models, each embedded with one of 14 distinct hidden behaviors, such as sycophantic deference or opposition to AI regulation. These models are specifically trained not to reveal their hidden behaviors when directly queried, providing a challenging environment for testing investigator agents and auditing tools.
Is Auditbench trustworthy and reputable?
StartupHub's Data Trust & Reputation score for Auditbench is 39 out of 100, based on site security posture and privacy practices.
When was Auditbench founded?
Auditbench was founded in 2026.
What industry does Auditbench operate in?
Auditbench operates in AI Safety, AI Alignment, Foundation Model, Large Language Model, Generative AI, AI Testing.
Your AI already decides. HONESTAS makes those decisions provable: deliberated by a chaired committee, cross-model gated, certificate-sealed and replayable.
Shisai checks whether AI-drafted text is backed by the records it cites, and produces evidence a compliance officer can file. Deterministic. No model in the judgement.
Open-source infrastructure for permissioned agent-to-agent communication: a deny-by-default kernel for file access and tool calls, with pluggable agent runtimes.
Dstilled follows AI companies, summarizing important announcements and filtering out noise to provide a ranked reading feed.
A technology licensing company specializing in identity-based cognitive architectures for generative systems, enabling AI to have a coherent self and grow through dialogue.
Aetios AI builds advanced clinical tools for mental health professionals, leveraging AI for psychological formulation and treatment planning.
Applied research notes on model internals, representation geometry, attention, and numerical precision.
White Sky develops institutional-grade embodied intelligence, autonomous mission systems, and deployment infrastructure for government, critical infrastructure, and enterprise environments.
An entrant is a company tagged AI Safety whose domain was first registered in the window, counted from registry records in the StartupHub directory. 13 of them registered in the last 30 days. Registry detection runs two to three weeks behind registration, so recent weeks are a floor.
No comments yet. Be the first to share your take.