Ory Adds Build and Deploy Tooling for AI Agent Security
Ory extends its identity and access platform with new build-time and deploy-time capabilities aimed at securing AI agents across the lifecycle.
AI Agents
Practical reporting on AI systems, tools, and engineering execution.
Git repos
Ory releases Build and Deploy capabilities to secure AI agents from development through production
Ory announces Build and Deploy capabilities that extend AI agent security across the development‑to‑production lifecycle, though performance details are not yet disclosed.
Sharding Git Quorum by Reference to Move Repositories to Azure
GitHub restructures its Spokes quorum from repository-wide locks to per-reference serialization, enabling the busiest repos to migrate to Azure without hitting a 25ms cross-region write ceiling.
Guardrails Aren’t Magic: What Happens When AI Safety Controls Meet Real-World Attacks
Most organizations assume their LLM guardrails block harmful prompts—but research shows adversarial inputs can bypass safety systems, so enterprise teams need layered defenses beyond text-filtering.
Cooperation Without Adaptation: LLM Agents in a Collective‑Risk Dilemma
In a multi‑round public‑goods test, five production LLMs often used near‑fixed contribution rules, failing to adapt to changing risk and partner behaviour — a caution for multi‑agent deployments and evaluation design.
BRIA: Isolate tool probes into branches, inspect results, adopt selectively — implementation notes for builders
BRIA isolates candidate tool investigations in temporary branches, inspects execution evidence, and adopts only useful results — a practical pattern for discovering relevant tools from very large libraries.
CDSE: Splitting LLM Agent Skills into Meta, Experience, Tool, and Script for Self-Improvement
CDSE decomposes a unified agent skill into four coordinated components and reports consistent score gains across 36 agent–harness–benchmark settings, highlighting functional heterogeneity in learned skills.
TIDER: RL-trained temporal windows to retrieve time-aligned memories for long-horizon agents
TIDER uses an RL-trained temporal-window locator plus a semantic channel to prioritize time-aligned memories, improving retrieval accuracy on long-horizon benchmarks.
Darwin-Claw: distill episodic experience into prototype vectors for controllable agent adaptation
Darwin-Claw proposes a Hierarchical Experience Consolidation system that distills episodic errors into prototype vectors to steer continual skill updates; here’s how to test and integrate it safely in production agents.
SkillBrowser: Meta‑Skill–managed on‑demand skill loading to reduce context pollution in LLM agents
SkillBrowser uses a Meta‑Skill manager to identify capability needs at runtime and load skills on demand, reporting aggregate accuracy gains over static retrieval across multiple benchmarks while adding latency and risk of missed skills.