audit

This paper discusses testing methodologies for financial controls in ERP data provisioning, likely aimed at improving audit and compliance processes.

AuditWeave is a lightweight Python library that records workflow steps into a tamper-evident, hash-chained ledger for auditing AI-assisted and data-transformation workflows, enabling traceability and integrity verification.

Allstate accuses Broadcom of launching retaliatory audits after it terminated its VMware and CA contracts, highlighting ongoing legal battles between Broadcom and disgruntled enterprise customers.

A tweet thread shares a comprehensive prompt for auditing vibe-coded projects before launch, covering security, reliability, concurrency, accessibility, and UI consistency.

An auditor finds that 0 of 32 published "discoveries" by an autonomous research agent were novel as framed, with the dominant failure being over-labeling rather than bad measurement.

This paper presents ElephantAgent, a protocol that enforces Contextual State Continuity to defend against contextual state poisoning in agentic systems, using replicated trusted hardware and historical traceability.

MemoryOps AI is an open-source memory governance layer for AI assistants that handles memory lifecycle with policies, expiration, auditing, and deletion guarantees. The author seeks technical feedback from developers building AI agents and RAG systems.

Introduces the Contagion Tensor and Coupling Amplification Factor (CAF) as a baseline-referenced, unitless ratio to quantify output-distribution coupling in multi-agent LLM systems, validated with real-API experiments on DeepSeek-Chat and GPT-4o-mini.

MetaviewAI used Devin AI agent to complete SOC2 audit in two days instead of weeks.

This paper proposes a governance model for autonomous AI agents based on institutional attestation, where actions are governed through independently attested evidence rather than monitoring agent reasoning. It formalizes this approach with a proof-of-concept implementation for high-risk actions like clinical prescribing and software deployment.

This paper introduces a fail-closed certification protocol to determine when a forecasting leaderboard winner can be reliably used as deployment-ready top-1 advice, given a fixed decision interface and deployed utility. It presents a locked native audit that prevents overclaiming by blocking apparent forecast/deployment winner inversions.

An analysis of how transparent Google's DiffusionGemma model release is, discussing the implications for AI safety and accountability.

agent-pd is an open-source logging and monitoring tool for Claude Code agents that records all tool and permission events and replays them through deterministic detectors to catch rule violations, without blocking any actions.

This paper identifies a failure mode in long-horizon research agents where optimizing an aggregate metric can select candidates that improve the headline number but break critical subgroups (inversion). It proposes a search-discipline protocol with an external control loop that audits candidates based on disaggregated behavior rather than the score.

This paper formalizes the concept of signed compression progress on a sealed audit as a reward that is Goodhart-resistant, proving that cumulative reward telescopes to genuine audit improvement and providing bounds for finite audit panels. It identifies failure modes and validates results with experiments.

Satya Nadella advocates treating AI agents as employees with identities, permissions, and audits, and discusses Microsoft's tools for managing them.

LlamaIndex launches granular bounding boxes in LlamaParse, enabling visual citations for every word in a document to allow human audit of exact numbers and figures.

Fort is a command-line tool that audits and fixes Mac security issues with a single command.

A discussion on how teams handle authentication and identity management for AI agents in production, covering service accounts, credential storage, approval workflows, and auditing.

This paper conducts a behavioral audit of seven open-weight and closed-source LLMs across four U.S. cities, finding that racial steering in housing recommendations is an emergent behavior of the model's interpretive license, varying by user identity and city context.