Vinoo Ganesh describes a retrieval architecture that separates an LLM's planning from the production of factual numbers. Documents are preprocessed into structured facts with stable identifiers; the model selects references, while code resolves values and performs arithmetic. Final outputs are checked so unresolved citations can be corrected or refused. The talk distinguishes three failure sources: retrieving the wrong document, choosing the wrong value within the right document, and performing unreliable arithmetic. A deterministic lookup addresses only part of that problem: an existing identifier can still refer to the wrong period, entity or extracted value, so source correctness and relevance need their own checks. The security discussion applies the same discipline to access. Tenant scope and user authorization must constrain both search and direct document reads, with stale indexes and caller-supplied identities treated as failure cases. This is an inspectable design pattern described by a vendor, not independent evidence of zero hallucinations or universal financial accuracy.
Grounded numbers: resolve source IDs and enforce access on every retrieval
Related research
More curated notes connected through AI Engineering and Model Evaluation.
Pacing model development in an era of cyber-critical capabilities
OpenAI says preliminary evidence that Astra may meet its Critical cybersecurity threshold led it to pause frontier reinforcement-learning work for two weeks and keep its largest planned run on hold. New safeguards include stronger workload and network isolation, continuous boundary testing, token-level monitoring that escalates suspicious tool activity, and broader alignment checks for deception, reward hacking, and unauthorized access.
OWASP Top 10 for Agentic Applications for 2026
OWASP's community guide organizes agentic-system risk into ten categories, including goal hijacking, tool misuse, identity and privilege abuse, memory poisoning, insecure inter-agent communication, cascading failures, and rogue-agent behavior. It provides a shared taxonomy and mitigation starting point rather than a certification checklist or evidence that a deployed system is secure.
A blueprint for democratic governance of frontier AI
OpenAI proposes a three-part U.S. frontier-AI governance model: harmonize emerging state safety laws into a federal baseline, strengthen CAISI as an evaluation and standards institution, and coordinate a broader resilience program. Proposed controls include severe-risk evaluations, transparency reports, independent audits, safety-incident reporting, model-weight security, whistleblower protection, and periodic technical assessments.