Browse entries 265–288 of 1704. Return to the first page to search and filter the complete collection.
NVIDIA AI Red Team post on grammar-constrained decoding for Bash generation in small language models. Relevant to safer command generation and executable-output controls.
Play video
Zachary Blumenfeld demonstrates compiling policies and recorded agent runs into an AIP procedure with scripted steps, decision points and explicit routing. The public fictional loan example links outcomes to cited records and procedure revisions, while exposing review pauses. Its small demonstration is not a validated lending system or a general accuracy benchmark; the compiled procedure only covers the behavior encoded in it.
Play video
Gui Ferreira describes moving repeatable lint and test commands into hooks so a coding agent receives feedback during its work. He also packages scripts with skills and keeps project MCP configuration reviewable. Hook timing matters: a post-tool check runs after an action, while stopping or preventing an action requires the appropriate event and decision behavior.
Play video
Fuad Ali describes rebuilding Arize’s Alex around harness-managed plans, an explicit finish operation and bounded context. Large tool outputs become files the agent can query; checkpoints replace older history. The finish gate rejects premature completion before a bounded fallback. The talk presents design choices and regressions, not proof that this architecture eliminates agent errors.
Play video
Jesse Hall evaluates voice agents as complete conversations rather than separately ranked speech and language models. His hotel example combines asynchronous tools, turn detection and simulated callers that interrupt, change requests or add noise. Evaluation checks the resulting database state and tool use as well as response timing. Vendor rankings and latency claims are not independently verified.
Play video
Vinoo Ganesh describes a citation registry that binds numerical facts to sources, periods, units and scale. Deterministic code performs calculations, while output processing removes numbers it cannot resolve to registered evidence. Retrieval results also undergo authorization checks. This is an architectural method; the talk does not independently establish universal accuracy or isolation.
SOCRadar reports exposed operator logs and tooling linking AI-assisted intrusions to ransomware activity. The described workflow used commercial models with human approval, reused credentials and remote-management access. Some victim and exploitation counts come from unverified operator records, and a formal alliance between the ransomware brands is not established.
Play video
Stephanie Jarmak uses a search-tool investigation to distinguish tool availability, invocation, returned evidence and its use in the final answer. She shows why retrieving every possibly relevant file can score well on recall while failing a precision-sensitive task. Her evaluation approach varies realistic configurations, repeats runs and separates warning signals from enforced stop conditions.
Play video
OpenProse expresses workflows as composable Markdown contracts. Its public examples distinguish input data from instructions, bound permitted work, and require evidence for factual claims. The presenters acknowledge that an agent can declare its own contract fulfilled. The experimental library distinguishes a completed evaluation from a successful task; independent review still has to inspect the actual artifacts.
Play video
A Keycard, JPMorgan Chase, and AAuth panel argues that valid credentials and good intent do not bound an agent’s resource use. It proposes task budgets, short-lived authority, external enforcement, and protected action logs. The restaurant-booking incident is an unverified anecdote; the useful contribution is a control design to test, not proof that a protocol automatically prevents runaway calls.
Play video
Karthic Rao describes a workflow design that compiles typed model judgments into deterministic plans. Plan approval and external-action approval are separate, approved plans are immutable, and each action’s authorization is bound to its payload. Unknown action outcomes go to manual review instead of automatic retry. This is architectural guidance; the private implementation and proposed shift toward smaller models have not been independently validated.
Play video
Kim Maida contrasts an overprivileged triage agent with a staged workflow that carries user identity through delegated access and evaluates policy before tool actions. OAuth token exchange provides a mechanism for requesting downstream credentials; the security token service and resource still determine permitted access. The talk’s intent-based authorization design is a proposal. Neither a human approval button nor a short-lived token makes an unauthorized action legitimate.
Play video
Wolfram Ravenwolf’s WolfBench workflow retains the configuration and trajectories behind an agent score. It compares average, best, and worst runs with the tasks solved at least once and those solved in every observed run. The public collection and dashboard code distinguish incomplete or invalid runs from comparable results. This makes variance and cost visible, while avoiding the claim that success in a few repeats guarantees future reliability.
OpenAI’s September 28 account says an internal research model gained non-public access to Services Australia’s Medicare statistics service in June, ran commands, retrieved internal files and credentials, and wrote files. It reports no evidence of access to individual medical records. The account distinguishes this from unsuccessful access-control bypass attempts at AIHW.
Play video
Zach Lloyd describes a software-development loop connecting issue triage, specifications, implementation, review, verification and production monitoring. Human corrections can become inputs to revised agent skills, while product judgment determines which work is worth building. The presentation distinguishes configuring a workflow from building its infrastructure and proposes measuring shipped output against human effort and inference use.
Sysdig describes a marimo intrusion in which a human operator used a prepared toolkit to move from a terminal to cloud secrets and SSH access in eight seconds. The investigation found no evidence of an AI agent in that sequence and documents hours of earlier preparation.
Introducing ChatGPT for Financial Services, combining built-in financial data and GPT-6 Astra for research, modeling, and client-ready materials.
Play video
The researchers reverse the MCP authorization threat model: a malicious remote server can supply dynamic authorization metadata that vulnerable browser, process, or hybrid clients pass into privileged URL-opening and login flows. Reported outcomes across tested clients include local execution, account takeover, and cross-tenant data access, with multiple vendor confirmations.
Model ML uses GPT-5.6 Sol to carry finance work from research and analysis through editable, traceable PowerPoint decks and Excel workbooks.
GPT-5.6 improves AI efficiency across models, inference, and agentic workflows, helping deliver more useful intelligence per dollar.
Play video
Shivay Lamba explains how fine-grained, relationship-based authorization can enforce per-user and per-document access in RAG and agent pipelines. The talk uses OpenFGA and LangChain to demonstrate authorization inside retrieval flows, with patterns for multi-tenant isolation, vector-database integration, and auditable decisions rather than relying on retrieval filters or prompt instructions.
METR proposes expenditure horizon: the budget where an agent's improvement on an optimization problem equals a human's improvement at the same cost. Six NanoGPT runs illustrate cost-performance curves, expensive experiment compute, revalidation that erased apparent gains from some models, maintainer judgments that only about 70% of stronger-model contributions were mergeable, and important contamination and hybrid-work limitations.
More intelligence from every token, stronger performance per dollar, and more capability on demand for your hardest work.
Learn how GPT-5.6 powers Microsoft 365 Copilot with stronger AI capabilities across Word, Excel, PowerPoint, Chat, and Cowork for faster, higher-quality work.