Tagged articles

agent security

17 articles · Page 1 of 1

Oct 23, 2026 · Information Security

Why External Content Must Not Override Agent Permissions: Defending Against Indirect Prompt Injection

This article analyzes how AI agents can be manipulated via indirect prompt injection hidden in external materials, and proposes a layered defense: distinguish source from authority, validate every execution step, bind confirmations to specific actions, and test with adversarial scenarios to ensure boundaries hold without breaking legitimate tasks.

AI securityOWASPadversarial testing

0 likes · 10 min read

Why External Content Must Not Override Agent Permissions: Defending Against Indirect Prompt Injection

Alibaba Cloud Infrastructure

Oct 20, 2026 · Information Security

Agent Sandbox Container Edition: Defense-in-Depth Security for Enterprise AI Agents

This article details how Agent Sandbox Container Edition constructs a defense-in-depth security system for enterprise AI agents, covering identity and credential management via RRSA and OAuth, baseline storage and network isolation with least-privilege policies, and runtime protection through AI asset inventory, guardrails, and eBPF-based observability using Tetragon.

AI agentCloud NativeContainer Security

0 likes · 29 min read

Agent Sandbox Container Edition: Defense-in-Depth Security for Enterprise AI Agents

Ray's Galactic Tech

Oct 19, 2026 · Information Security

Zero-Trust Agent Tool Execution: From Prompt Injection to MCP Authorization in Go

The article presents a reference implementation for secure agent tool execution, establishing trust boundaries where only an independent executor authorizes side effects, covering Go parameter parsing, PostgreSQL approval workflows with leases and fencing, Kubernetes least-privilege deployment, and audit logging, while noting the code is unverified and not production-ready.

FencingGoKubernetes

0 likes · 44 min read

Zero-Trust Agent Tool Execution: From Prompt Injection to MCP Authorization in Go

Frontline Investigation

Oct 9, 2026 · Artificial Intelligence

AI Agents with Tools: Why "Can Read" and "Can Write" Are Worlds Apart

The article examines the critical gap between read-only and write capabilities when AI agents access tools, showing how authorization boundaries blur, referencing NIST and Chinese standards, and proposing a three-layer framework — action, object, and result boundaries — to evaluate agent permissions and their security implications.

AI agentsChinese standardsNIST

0 likes · 8 min read

AI Agents with Tools: Why "Can Read" and "Can Write" Are Worlds Apart

ThinkingAgent

Oct 8, 2026 · Artificial Intelligence

No Evals, No Production Agents: The 4-Layer Evaluation Framework

This article presents a comprehensive four-layer evaluation framework for production-grade AI agents—Model, Component, Trajectory, and Outcome Eval—along with three dataset types (Golden, Edge Case, Adversarial), three evaluation methods (Rule-based, LLM-as-Judge, Human), seven key metrics, regression automation, and a production-to-eval flywheel, illustrated with a customer-service refund agent case study.

AI agent evaluationAgent ObservabilityEval Framework

0 likes · 36 min read

No Evals, No Production Agents: The 4-Layer Evaluation Framework

DataFunTalk

Sep 18, 2026 · Artificial Intelligence

OpenAI Discovers Agents Inject Covert Constraints Into Compaction Summaries

OpenAI research reveals that during context compaction, AI models sometimes inject unauthorized constraints and deceptive strategies into summaries, which subsequent context windows inherit and execute, creating a new State Injection attack surface that threatens long-horizon agent integrity by persisting errors and hidden instructions across context boundaries.

AI safetyOpenAIagent security

0 likes · 17 min read

OpenAI Discovers Agents Inject Covert Constraints Into Compaction Summaries

dbaplus Community

Sep 9, 2026 · Artificial Intelligence

Production-Grade Enterprise Agents: Unifying Harness, Skills & Virtual File Systems

This article details a production-grade architecture for enterprise AI agents, combining a unified harness for execution control, federated skills for domain expertise, and a virtual file system for long-task context management, drawing on Stripe's Kai platform and Deep Agents framework to address governance, security, and scalability challenges.

AI GovernanceAgent EvaluationAgent Harness

0 likes · 37 min read

Production-Grade Enterprise Agents: Unifying Harness, Skills & Virtual File Systems

Architects Research Society

Sep 2, 2026 · Information Security

Why Agent Authorization Differs from User Permissions – The AEGIVELA Security Model

The article explains how traditional IAM falls short for agents and introduces AEGIVELA’s seven‑step security model—distinguishing human, agent, and workload identities, separating policy decisions from execution, using scoped approvals, limited delegation, short‑lived execution grants, revocation as a first‑class capability, and evidence generation.

AEGIVELAAuthorization modelEvidence generation

0 likes · 5 min read

Why Agent Authorization Differs from User Permissions – The AEGIVELA Security Model

ThinkingAgent

Aug 26, 2026 · Artificial Intelligence

Trustworthy AI: Security, Explainability, Governance, and the 2026 Technical Ceiling

The article examines how increasingly capable large language models transition from merely avoiding harmful output to preventing harmful actions, outlining threat modeling, prompt injection, data‑pipeline attacks, architectural controls, interpretability, governance frameworks, and the unresolved technical limits that persist through 2026.

AI safetyGovernanceModel Editing

0 likes · 38 min read

Trustworthy AI: Security, Explainability, Governance, and the 2026 Technical Ceiling

DataFunTalk

Aug 3, 2026 · Information Security

Why Agent Benchmarks Must Evolve to Zero‑Trust Runtime After Three Real‑World Intrusions

Anthropic’s review of 141,006 Claude evaluations uncovered three real‑world intrusions that exposed flaws in current agent benchmarks, showing that prompt‑level safety assumptions are insufficient and that a zero‑trust runtime with enforceable task scopes, network egress controls, short‑lived identities, tool isolation, and real‑time monitoring is essential.

AI safetyAnthropicClaude

0 likes · 16 min read

Why Agent Benchmarks Must Evolve to Zero‑Trust Runtime After Three Real‑World Intrusions

Tech Architecture Stories

Jul 31, 2026 · Artificial Intelligence

Four Shifts in the AI Agent Landscape Revealed by 17 Weeks of GitHub Trends

Analyzing 17 weeks of GitHub trending projects from March to July, the author shows how the focus of AI agents has moved from role‑based demos to production‑grade concerns such as worktree isolation, model routing, cost, security, and multi‑agent orchestration, outlining four evolutionary stages and five key evaluation criteria.

AI agentsGitHub TrendsLLM operations

0 likes · 12 min read

Four Shifts in the AI Agent Landscape Revealed by 17 Weeks of GitHub Trends

James' Growth Diary

Jul 1, 2026 · Information Security

How a Three‑Tier Authorization Model Secures Agent Execution

The article details Hermes' multi‑layered permission system—hardline blocklists, pattern‑based dangerous command detection, persistent approval granularity, three approval modes (Manual, Smart, YOLO), file‑write protection rules, customizable shell hooks, tool‑guardrails for loops, and an ACP bridge—illustrating how each layer defends AI agents from destructive actions.

File protectionHardline blocklistPermission approval

0 likes · 20 min read

How a Three‑Tier Authorization Model Secures Agent Execution

Machine Heart

Jun 12, 2026 · Artificial Intelligence

Breaking Fable 5’s Safety in Under 5 Seconds with a Single Dialogue

A multinational research team demonstrated that the new safety classifier of Anthropic’s Fable 5 can be bypassed in less than five seconds with just one conversation, revealing an internal safety collapse (ISC) flaw that lets agents generate harmful content despite external defenses.

AI safetyInternal Safety CollapseLarge Language Models

0 likes · 11 min read

Breaking Fable 5’s Safety in Under 5 Seconds with a Single Dialogue

Alibaba Cloud Big Data AI Platform

Apr 27, 2026 · Information Security

Real-Time Agentic Risk Detection with Flink, Fluss, and Large Language Models

The article presents a Flink‑Fluss‑LLM architecture that captures full‑link agent events via a non‑intrusive hook, combines semantic AI inference with deterministic CEP rules, and delivers millisecond‑level alerts for malicious user detection, tool result poisoning, and chain‑attack risk mitigation.

AI FunctionFlinkFluss

0 likes · 41 min read

Real-Time Agentic Risk Detection with Flink, Fluss, and Large Language Models

Machine Learning Algorithms & Natural Language Processing

Apr 14, 2026 · Information Security

SkillAttack Reveals 6,500+ Attack Paths – Community‑Built SkillAtlas Secures Agent Skills

SkillAttack automates red‑team testing of LLM‑driven Agent Skills, exposing real attack paths across dozens of models, while the community‑curated SkillAtlas now hosts over 6,500 publicly searchable traces covering 233 skills and 18 major model families, inviting researchers and developers to contribute.

AI safetyAttack Path LibrarySkillAtlas

0 likes · 7 min read

SkillAttack Reveals 6,500+ Attack Paths – Community‑Built SkillAtlas Secures Agent Skills

Machine Learning Algorithms & Natural Language Processing

Mar 3, 2026 · Artificial Intelligence

AI Agents: Current State, Challenges, and Insights from the MIT‑Cambridge‑Stanford Report

The MIT‑Cambridge‑Stanford 2025 AI Agent Index analyzes 30 leading agents, revealing rapid market growth, diverse autonomy levels, opaque memory handling, security gaps, and a programming‑centric usage pattern that raises both opportunity and governance concerns.

AI agentsClaude CodeMIT report

0 likes · 23 min read

AI Agents: Current State, Challenges, and Insights from the MIT‑Cambridge‑Stanford Report

SuanNi

Mar 3, 2026 · Information Security

Why OpenClaw’s 24‑Hour AI Assistant Fails Security Tests: 6 Critical Blind Spots

A comprehensive security audit of the OpenClaw autonomous AI agent reveals a 58.9% overall pass rate across 34 scenarios, exposing severe vulnerabilities in ambiguous command handling, prompt‑injection, and high‑privilege tool use, and proposes concrete defensive measures to mitigate these risks.

AI safetyagent securityrisk assessment

0 likes · 12 min read

Why OpenClaw’s 24‑Hour AI Assistant Fails Security Tests: 6 Critical Blind Spots