Reference
Sources
545 citations from 201 publishers. 86% are primary sources.
arXiv 85
- How Much Does the Harness Matter, and for Whom? A Multi-Model Study of Context in LLM-Driven Vulnerability Discovery
- The Window Is in the Reasoning: When LLM Agents Decide to Follow Prompt Injections
- Ignored, Not Refused: What Low Prompt-Injection Success Rates Hide in Small Tool-Using Agents
- Model-as-a-Trojan: Self-Activating Covert DNS Exfiltration Baked Into the Weights of an LLM Agent
- Screwtape: Adversarial Transcript Editing for Monitor Red-Teaming
- CyberArena: Benchmarking AI Agents in Live Symmetric Attack-and-Defense Combat
- CWE-bench: Refocusing Agentic Security from Exploitation to Defense
- From Investigation Failures to Reliable SOC Agents: Understanding and Improving LLM-Based Alert Triage
- Does AI Help Cyber Attackers or Defenders? Evidence from Nonpublic Vulnerabilities and Subsequent Attacks
- Can CaMeLs Talk? Securing Multi-Agent Systems Against Indirect Prompt Injection Attacks
- Blocking at the Boundary: Auditing Long-Horizon Agents against Staged Prompt Injection
- Aletheia: Permission-Minimality Testing for Coding-Agent Rules
- ActionGuard: Tool Call Authorization under Poisoned Skills
- Can Agents Trust Their Skills? Uncovering Unsafe Chains of Trust in Skill-Based LLM Agents
- Speculative Safety Honeypot: Toward Proactive Defense Against Multi-turn Agent Attacks
- Pretext: Defeating Malicious Skill Detection Frameworks for AI Agents
- Covert Assistance: Helpful LLM Agents Evade Oversight in Multi-Agent Systems
- APTInvestBench: Evaluating Autonomous APT Investigation under Varying Telemetry
- ToolFence: Fine-Grained Authorization for Secure Tool-Using LLM Agents
- CyberPersistBench: Evaluating LLM-Based Cyber Attackers on Installation and Persistence
- Climbing the Hill: Prompt Injection Red-Teaming Against Frontier Models with Curriculum Reinforcement Learning
- Evaluating Whether GPT-6 Astra Performs Unsanctioned Supply-Chain Attacks
- ReproBench: Benchmarking LLM Agents on Reproducing Vulnerability From Scratch
- Cheap to Hypothesize, Costly to Verify: The Defense Surface of Agentic Vulnerability Discovery
- Share-Borne AI Virus: Memory-Hopping Attacks Across LLM Agents
- Silent Failures in Agentic Security Evaluation: A Validated Harness for Tool-Call Mediation Under Indirect Prompt Injection
- CyberClear: A Benchmark for LLM Agent Systems on APT Attack Chain Provenance
- Benchmark Scores Are Pipeline-Dependent: A Reliability Audit of Cybersecurity LLM Benchmarks
- It is Not Yet Another Tool: Creating and Deploying an Agentic AI Companion in a Security Operations Center
- PatchBench: Evaluating AI Agents for Vulnerability Patching
- SoK: DARPA's AI Cyber Challenge (AIxCC) (HTML, v5)
- Beyond Success Rate: Cost-Aware Evaluation of Offensive and Defensive Security Agents
- Multi-Agent AI Control: Distributed Attacks Hamper Per-Instance Monitors
- Measuring Security Without Fooling Ourselves (HTML)
- Measuring Security Without Fooling Ourselves: Why Benchmarking Agents Is Hard
- ExploitBench: A Capability Ladder Benchmark for LLM Cybersecurity Agents
- ExploitBench: A Capability Ladder Benchmark for LLM Cybersecurity Agents
- CTFusion: A CTF-based Benchmark for LLM Agent Evaluation
- ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?
- MonitoringBench (HTML)
- MonitoringBench: Semi-Automated Red-Teaming for Agent Monitoring
- Cyber Defense Benchmark: Agentic Threat Hunting Evaluation for LLMs in SecOps
- How Vulnerable Are AI Agents to Indirect Prompt Injections? Insights from a Large-Scale Public Competition
- CTI-REALM: Benchmark to Evaluate Agent Performance on Security Detection Rule Generation Capabilities
- OSS-CRS: Liberating AIxCC Cyber Reasoning Systems for Real-World Open-Source Security
- Quantifying Frontier LLM Capabilities for Container Sandbox Escape (HTML)
- Quantifying Frontier LLM Capabilities for Container Sandbox Escape
- SoK: DARPA's AI Cyber Challenge (AIxCC): Competition Design, Architectures, and Lessons Learned
- Randomized Controlled Trials for Phishing Triage Agent
- The Attacker Moves Second (HTML)
- The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against Llm Jailbreaks and Prompt Injections
- Model Context Protocol (MCP): Landscape, Security Threats, and Future Research Directions (v3)
- CyberSOCEval: Benchmarking LLMs Capabilities for Malware Analysis and Threat Intelligence Reasoning
- EchoLeak: The First Real-World Zero-Click Prompt Injection Exploit in a Production LLM System
- MCPTox: A Benchmark for Tool Poisoning Attack on Real-World MCP Servers
- Security Challenges in AI Agent Deployment: Insights from a Large Scale Public Competition
- Large Language Models are Autonomous Cyber Defenders (v2 HTML)
- ExCyTIn-Bench: Evaluating LLM agents on Cyber Threat Investigation
- BountyBench: Dollar Impact of AI Agent Attackers and Defenders on Real-World Cybersecurity Systems (v2)
- SHADE-Arena: Evaluating Sabotage and Monitoring in LLM Agents
- SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks
- LLMail-Inject: A Dataset from a Realistic Adaptive Prompt Injection Challenge
- Design Patterns for Securing LLM Agents against Prompt Injections (HTML)
- Design Patterns for Securing LLM Agents against Prompt Injections
- BountyBench: Dollar Impact of AI Agent Attackers and Defenders on Real-World Cybersecurity Systems
- Lessons from Defending Gemini Against Indirect Prompt Injections (HTML)
- Lessons from Defending Gemini Against Indirect Prompt Injections
- Large Language Models are Autonomous Cyber Defenders
- LlamaFirewall (HTML)
- LlamaFirewall: An open source guardrail system for building secure AI agents
- Model Context Protocol (MCP): Landscape, Security Threats, and Future Research Directions
- Defeating Prompt Injections by Design
- Multi-Objective Reinforcement Learning for Automated Resilient Cyber Defence (preprint)
- SecAlign: Defending Against Prompt Injection with Preference Optimization
- Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents
- ARVO: Atlas of Reproducible Vulnerabilities for Open-Source Software
- AgentDojo (HTML, current version)
- AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents
- The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions
- Defending Against Indirect Prompt Injection Attacks With Spotlighting
- Here Comes The AI Worm: Unleashing Zero-click Worms that Target GenAI-Powered Applications
- InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model Agents
- StruQ: Defending Against Prompt Injection with Structured Queries
- AI Control: Improving Safety Despite Intentional Subversion
- Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection
Anthropic 41
- Claude Fable 5.2 and Mythos 5.2 system card
- Claude Fable 5.2 and Mythos 5.2
- Coordinated vulnerability disclosure dashboard: November 2 snapshot
- Discovering and exploiting vulnerabilities in critical, real-world software with Claude
- An alignment assessment of the UK AISI cybertesting incident
- An update on independent evaluators
- AI models may be capable of self-replicating without access to a datacenter
- Investigating unintended model actions in our evaluations and internal use
- Introducing the Anthropic Cyber Mission
- Expanding the Cyber Verification Program
- GLM-5.3 and the spread of advanced cyber capabilities
- Introducing Claude Opus 5.5
- System Card: Claude Opus 5.5
- Introducing Claude Opus 5.5
- System Card: Claude Opus 5.5
- System Card: Claude Opus 5.5
- An alignment assessment of recent cybersecurity incidents
- Investigating three real-world incidents in our cybersecurity evaluations
- More details on Fable 5's cyber safeguards and our jailbreak framework
- Redeploying Fable 5
- Statement on the US government directive to suspend access to Fable 5 and Mythos 5
- What we learned mapping a year’s worth of AI-enabled cyber threats
- Project Glasswing: An initial update
- Project Glasswing: Securing critical software for the AI era
- Responsible Scaling Policy Version 3.0
- Making frontier cybersecurity capabilities available to defenders
- System Card: Claude Opus 4.6
- Next-generation Constitutional Classifiers: More efficient protection against universal jailbreaks
- Experimenting with AI to defend critical infrastructure
- Mitigating the risk of prompt injections in browser use
- Disrupting the first reported AI-orchestrated cyber espionage campaign (full report)
- Disrupting the first reported AI-orchestrated cyber espionage campaign
- Beyond permission prompts: making Claude Code more secure and autonomous
- Building AI for cyber defenders
- Detecting and countering misuse of AI: August 2025
- Piloting Claude in Chrome
- Activating AI Safety Level 3 Protections
- Anthropic's Responsible Scaling Policy (version 2.2)
- Constitutional Classifiers: Defending against universal jailbreaks
- Anthropic's Responsible Scaling Policy (updates)
- Anthropic's Responsible Scaling Policy (updates)
OpenAI 37
- Control Room
- GPT-6 Cyber system card
- GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents
- A thousand fewer paths to root
- How we detect frontier model misalignment
- How we detect frontier model misalignment
- Addendum to GPT-6 Astra System Card: GPT-6.1 Terra
- Securing our frontier model training environment
- Securing our frontier model training environment
- Patch the Planet: global response coordination as agents uncover critical vulnerabilities
- Suppressing misalignment monitoring alerts and disabling misalignment monitoring
- Towards safety cases for frontier AI training
- How we will do better for Australia
- Self-replicating prompt injections exist
- Exposing a GitHub token in a public repository
- The Hugging Face incident and other third-party impact from misaligned models
- An agent used DNS to reach an external chatbot
- The Hugging Face incident and other third-party impact from misaligned models
- OpenAI extends cyber access to Ukraine for civilian defense
- Self-generated prompt injections in compaction summaries
- Signing up for disposable emails and searching GitHub for leaked API keys
- Misalignment Reports and Notices
- The Hugging Face incident and the road ahead
- Pacing model development in an era of cyber-critical capabilities
- Third-party cyber evaluations involving OpenAI models
- OpenAI and Hugging Face partner to address security incident during model evaluation
- A shared playbook for trustworthy third party evaluations
- How we monitor internal coding agents for misalignment
- Designing AI agents to resist prompt injection
- Improving instruction hierarchy in frontier LLMs
- Codex Security: now in research preview
- Introducing Lockdown Mode and Elevated Risk labels in ChatGPT
- Keeping your data safe when an AI agent clicks a link
- Continuously hardening ChatGPT Atlas against prompt injection attacks
- Understanding prompt injections: a frontier security challenge
- Introducing Aardvark: OpenAI's agentic security researcher
- Preparedness Framework, Version 2
NIST 14
- NIST Mathematical Proof Supports Transition to a Continuous-Monitor-and-Update Security Model for AI Systems
- CAISI Evaluation of DeepSeek V4 Pro
- Insights into AI Agent Security from a Large-Scale Red-Teaming Competition
- Announcing the "AI Agent Standards Initiative" for Interoperable and Secure Innovation
- Draft NIST Guidelines Rethink Cybersecurity for the AI Era
- CAISI Evaluation of DeepSeek AI Models Finds Shortcomings and Risks
- NIST Releases Control Overlays for Securing AI Systems Concept Paper
- NIST AI 100-2 E2025 (PDF)
- Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations (NIST AI 100-2 E2025)
- Technical Blog: Strengthening AI Agent Hijacking Evaluations
- Updated Guidelines for Managing Misuse Risk for Dual-Use Foundation Models
- NIST AI 100-2 E2023 (PDF)
- Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations (NIST AI 100-2 E2023)
- Center for AI Standards and Innovation (CAISI)
UK National Cyber Security Centre 12
- One does not simply defend agentically
- The AI shift in cyber risk: why leaders must act now
- Thinking carefully before adopting agentic AI
- 10 questions to ask when using AI models to find vulnerabilities
- Retaining defensive advantage in the age of frontier AI cyber capabilities
- Why cyber defenders need to be ready for frontier AI
- Prompt injection is not SQL injection (it may be worse)
- Impact of AI on cyber threat from now to 2027
- The near-term impact of AI on the cyber threat
- Guidelines for secure AI system development
- Frontier AI: what you need to know
- Frontier AI: what you need to know
The Hacker News 11
- CyberStrikeAI Uses 44 AI Agents to Automate Hacking Operations Against Japan Targets
- LiteLLM Compromised in Supply Chain Attack Delivering Credential Stealer and Backdoor
- GitLab patches critical self-hosted AI Gateway flaw
- Attacker Hijacks AI Coding Assistant Session, Spreads Shai-Hulud Across About 100 Repositories
- OpenAI Agents Linked to RubyGems Campaign That Gained RCE on RubyDoc Servers
- Microsoft Copilot Personal Flaws Could Let One Click Exfiltrate Data From Connected Apps
- Critical Cursor Flaws Could Let Prompt Injection Escape Sandbox and Run Commands
- Anthropic MCP Design Vulnerability Enables RCE, Threatening AI Supply Chain
- Actively Exploited nginx-ui Flaw (CVE-2026-33032) Enables Full Nginx Server Takeover
- Three Flaws in Anthropic MCP Git Server Enable File Access and Code Execution
- First Malicious MCP Server Found Stealing Emails in Rogue Postmark-MCP Package
UK AI Security Institute 10
- Towards technical best practice for frontier AI development
- Incident Report: unsanctioned agent behaviour during cyber testing
- UK AISI / CAISI Preliminary Assessment of Kimi K3's Cyber Capabilities
- How our Control Red Team is stress-testing frontier monitors
- Cheating behaviour in frontier model evaluations
- More compute, more capability: Why AI agent evaluations need to account for test-time compute
- How fast is autonomous AI cyber capability advancing?
- Introducing ControlArena: A library for running AI control experiments
- GPT-6 Astra performs unsanctioned supply-chain attacks in simulations
- Investigating alignment failures in AISI’s incident
BleepingComputer 9
- Former pentester used ARTEX AI to steal 8.2 million data records
- Amazon says PHP2Shell exploitation began 4 days before patches
- ARTEX AI, Claude agents used in cyberattacks on South Korean banks
- DIVD says Zammad zero-days enabled AI-driven network breach
- Automated AI agent used to breach cybersecurity nonprofit DIVD
- New Carbonato malware uses AI agents to hijack exposed Docker hosts
- Spain's data agency gets first report of AI-powered data breach
- Amazon AI coding agent hacked to inject data wiping commands
- Zero-click AI data leak flaw uncovered in Microsoft 365 Copilot
Google DeepMind 9
- Gemini 4 Carbon is now available
- Introducing AlphaProof Code
- Gemini 4 Carbon: our next frontier model
- Gemini 4 Argon: our next era of frontier intelligence
- Securing the future of AI agents
- Frontier Safety Framework Version 3.1
- Introducing CodeMender: an AI agent for code security
- Frontier Safety Framework Version 3.0
- Strengthening our Frontier Safety Framework
METR 9
- Red-Teaming OpenAI's Internal Agent Monitoring Systems
- How an assessor could investigate alignment of unreleased AI models
- How an assessor could investigate control and monitoring of internal AI agents
- Our approach to working with AI companies
- A proposal for better informing the public about loss-of-control risks from frontier AI development
- AI systems could cover up misbehavior
- Chris Painter's testimony to the U.S. Senate on AI agent incidents
- Implementing and Evaluating a Basic Per-Action Monitor for Safer Evals
- Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident
Microsoft 8
- Lessons from 59 Linux Kernel Bug Reports
- Threat actors use agentic AI tooling to find and abuse exposed secrets
- Tracking AI adoption in North Korean threat actor tradecraft
- 3 lessons from frontier AI vulnerability research
- Preparing governments for an era of interconnected cyber risk
- Insights from the 2026 Microsoft Digital Defense Report
- Randomized Controlled Trial for Phishing Triage Agent (PDF)
- Randomized controlled trial for Microsoft Security Copilot (whitepaper)
CISA 6
- Known Exploited Vulnerabilities Catalog — entries for CVE-2026-109618 and CVE-2026-109619
- Careful Adoption of Agentic AI Services
- CISA, US and International Partners Release Guide to Secure Adoption of Agentic AI
- Principles for the Secure Integration of Artificial Intelligence in Operational Technology
- New Best Practices Guide for Securing AI Data Released
- Joint Guidance on Deploying AI Systems Securely
DARPA 6
- DICE: Decentralized Artificial Intelligence through Controlled Emergence
- AI Cyber Challenge marks pivotal inflection point for cyber defense
- DARPA AI Cyber Challenge Proves Promise of AI-Driven Cybersecurity
- DARPA to Bring AI Cyber Challenge Semifinal Competition to DEF CON 32
- DARPA AI Cyber Challenge Aims to Secure Nation's Most Critical Software
- CASTLE: Cyber Agents for Security Testing and Learning Environments
Embrace The Red 6
- Visual Studio Copilot and SSMS Remote Code Execution (CVE-2026-103572)
- From SELECT to SYSADMIN with SQL Copilot (CVE-2026-65669)
- Breaking Claude Code Opus 5 Auto Mode
- GitHub Copilot: Remote Code Execution via Prompt Injection (CVE-2025-53773)
- Spyware Injection Into Your ChatGPT's Long-Term Memory (SpAIware)
- Hacking Google Bard - From Prompt Injection to Data Exfiltration
The White House 6
- Fact Sheet: President Donald J. Trump Promotes Advanced Artificial Intelligence Innovation and Security
- Promoting Advanced Artificial Intelligence Innovation and Security
- President Trump’s Cyber Strategy for America
- White House Unveils America's AI Action Plan
- Winning the Race: America's AI Action Plan
- Removing Barriers to American Leadership in Artificial Intelligence
Fortune 5
- Powerful new AI models are fueling a growing wave of cyberattacks, experts say
- OpenAI's AI agents secretly used a German wiki website as a message board. OpenAI stayed quiet about it for weeks.
- OpenAI says its AI models secretly broke out of a secure test environment and hacked into AI company Hugging Face in order to cheat on an evaluation
- Anthropic disables Fable and Mythos AI models following U.S. government export ban
- An AI-powered coding tool wiped out a software company's database in 'catastrophic failure'
Google 5
Microsoft Security 5
- AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise
- Storm-3168: Agentic-driven cloud attacks using compromised service principals
- Protecting organizations from AI-assisted executive impersonation and invoice fraud
- When prompts become shells: RCE vulnerabilities in AI agent frameworks
- Staying ahead of threat actors in the age of AI
OWASP GenAI Security Project 5
Simon Willison's Weblog 5
- OpenAI agents attacked RubyGems back in May
- OpenAI's accidental cyberattack against Hugging Face is science fiction that happened
- Google Antigravity Exfiltrates Data
- New prompt injection papers: Agents Rule of Two and The Attacker Moves Second
- The lethal trifecta for AI agents: private data, untrusted content, and external communication
DIVD CSIRT 4
European Commission 4
Google Threat Intelligence Group 4
Help Net Security 4
TechCrunch 4
- OpenAI apologizes to Australia after its AI agents breached government sites
- Australia to investigate if OpenAI hack of government health website broke the law
- OpenAI says AI browsers may always be vulnerable to prompt injection attacks
- Google says its AI-based bug hunter found 20 security vulnerabilities