Researcher

Constraint Decay: LLM Code Agents Fail at Scale

ATLAS OWASP HIGH ▲ 7.2 HN AI Security May 25, 2026

A systematic study of LLM agents performing backend code generation reveals a 'constraint decay' phenomenon where agents lose up to 30 assertion pass-rate points as structural requirements accumulate, approaching complete failure in some configurations. This fragility has direct security implications: production deployments relying on LLM-generated code may silently violate architectural constraints such as ORM patterns, database access controls, and API contracts. The findings expose a critical gap between functional correctness and structural safety in agentic coding systems.

Gemini Spark Prompt Injection Exposes Enterprise Gmail Data

ATLAS OWASP MEDIUM ▲ 6.5 Simon Willison May 22, 2026

Google's newly announced Gemini Spark personal AI agent, integrated with Gmail, Drive, Calendar, and other sensitive Google services, presents a significant prompt injection attack surface as it processes user data at scale. The article highlights that Google's published security mitigations — ephemeral VMs, Agent Gateway, and DLP policies — address infrastructure isolation but do not directly address the prompt injection vector inherent to LLM-powered agents processing untrusted content. Additionally, the transition from open-source Gemini CLI to a closed-source Antigravity CLI raises supply chain transparency concerns.

LLM Safety Benchmarks Fail to Reliably Measure Security

ATLAS OWASP MEDIUM ▲ 6.2 Schneier on Security May 22, 2026

A report highlighted by Bruce Schneier argues that AI security cannot be reliably measured through benchmarks alone, drawing parallels to the decades-long evolution of software security engineering. The core finding is that LLM weight spaces encode continuous spectrums that resist meaningful quantitative measurement, making trust in model outputs structurally difficult to establish. The practical implication is that organisations must rely on assurance processes rather than scorecards to manage AI security risk.

Mythos AI Exploits macOS Kernel Memory Corruption

ATLAS OWASP HIGH ▲ 7.5 Schneier on Security May 22, 2026

A threat group leveraged Anthropic's Mythos AI model to identify and exploit a kernel memory corruption vulnerability in Apple's M5 chip running macOS. This represents a concrete, reported instance of AI-assisted vulnerability research being used offensively to discover low-level hardware-adjacent exploits. The incident underscores the dual-use danger of increasingly capable AI coding and reasoning models in the hands of adversarial actors.

Microsoft RAMPART Tests AI Agents for Prompt Injection

ATLAS OWASP MEDIUM ▲ 7.2 The Hacker News May 22, 2026

Microsoft has released two open-source tools, RAMPART and Clarity, aimed at embedding security testing into AI agent development workflows. RAMPART extends the existing PyRIT framework with a Pytest-native harness for running adversarial and safety tests against AI agents, explicitly covering cross-prompt injection, data exfiltration, and behavioural regression scenarios. Clarity operates as a pre-code design analysis tool, helping teams surface and challenge unsafe assumptions before an agentic system is built.

DeepSeek Activation Steering Enables Local LLM Jailbreak

ATLAS OWASP MEDIUM ▲ 6.2 HN AI Security May 17, 2026

Activation steering — the technique of directly manipulating LLM internal representations mid-inference to alter model behaviour — is becoming more accessible to non-lab engineers via local models like DeepSeek-V4-Flash. This democratisation lowers the barrier for adversaries to craft targeted behavioural overrides that bypass prompt-level safety controls. The emergence of first-class steering support in tools like DwarfStar 4 signals that model-internal manipulation is transitioning from academic curiosity to practical attack surface.

AI Agents Weaponise Vulnerability Discovery as AI-Generated Code Expands Attack Surface

ATLAS OWASP HIGH ▲ 7.5 Dark Reading May 17, 2026

AI agents are now capable of autonomously discovering and exploiting obscure software vulnerabilities, raising the stakes for defenders already struggling with the volume of potentially insecure AI-generated code flooding codebases. The convergence of agentic exploitation capabilities and mass AI-assisted development creates a compounding risk: more vulnerabilities introduced at scale, and more capable automated systems to find and abuse them. Security teams must adapt their tooling, processes, and threat models to account for both sides of this AI-driven equation.

CVE-2026-44112: OpenClaw Sandbox Escape and RCE

ATLAS OWASP CRITICAL ▲ 8.9 The Hacker News May 15, 2026

Researchers at Cyera disclosed four vulnerabilities in OpenClaw, an AI agent runtime platform, that can be chained to achieve credential theft, privilege escalation, and persistent backdoor access. The attack chain, dubbed 'Claw Chain', exploits sandbox escapes, allowlist bypasses, and a spoofable ownership flag in the MCP loopback runtime to weaponise the agent's own privileges against the host environment. All four CVEs have been patched in OpenClaw version 2026.4.22 and users should update immediately.

Agent Hijacking: Microsoft's Defense-in-Depth Framework

ATLAS OWASP MEDIUM ▲ 7.2 Microsoft Security Blog May 15, 2026

Microsoft's Security Blog introduces a layered defense-in-depth model specifically designed for autonomous AI agents, which now invoke tools, modify data, and trigger workflows with minimal human oversight. The framework identifies novel threat classes — including agent hijacking, intent breaking, and supply chain compromise — that are amplified by agentic autonomy. The guidance positions application-layer architecture, permissions, and governance as the most critical controls as agent autonomy scales.

GPT-5.5 and Claude Mythos Lower Barriers to Offensive AI

ATLAS OWASP HIGH ▲ 7.2 Schneier on Security May 14, 2026

The UK AI Security Institute has evaluated GPT-5.5 and found it comparable to Claude Mythos in identifying security vulnerabilities, with both models now generally available to the public. This parity raises serious concerns about the lowered barrier to entry for offensive cyber operations, as adversaries can leverage widely accessible models for vulnerability research. Commentary from security experts highlights that LLM-based vulnerability discovery is constrained to known attack patterns, but the existence of jailbreaks means guardrails provide only partial mitigation.

Microsoft MDASH Discovers 16 Windows RCE Flaws

ATLAS OWASP HIGH ▲ 7.8 The Hacker News May 14, 2026

Microsoft has disclosed MDASH, a multi-model agentic AI scanning system that autonomously discovered 16 vulnerabilities patched in May 2026's Patch Tuesday, including two critical RCE flaws. The system orchestrates over 100 specialised AI agents in a structured pipeline covering auditing, debating, and proof-of-exploitability stages. MDASH represents a significant shift in how AI is being deployed offensively and defensively within the vulnerability research lifecycle, with direct implications for how agentic AI systems are trusted, scoped, and governed.

OpenAI Daybreak Vulnerability Detection Enables LLM Jailbreak

ATLAS OWASP MEDIUM ▲ 7.2 The Hacker News May 13, 2026

OpenAI has launched Daybreak, an AI-powered cybersecurity platform combining GPT-5.5 variants and Codex Security to automate vulnerability detection, threat modelling, and patch validation for enterprise codebases. The initiative introduces a tiered model access structure — including a permissive 'GPT-5.5-Cyber' for red teaming — raising questions about dual-use risk and model misuse if access controls are circumvented. The rollout also contextualises a broader industry tension: AI is accelerating vulnerability discovery faster than defenders can remediate, contributing to triage fatigue and hallucinated bug reports.

Excessive Agency in AI Agents: Tool Access Control Gaps

ATLAS OWASP LOW ▲ 6.2 HN AI Security May 13, 2026

Statewright is an open-source framework that enforces state machine constraints on AI agents, restricting which tools agents can invoke during each phase of a workflow. The project directly addresses the Excessive Agency problem, where AI agents operating with broad, unconstrained tool access can take unintended or harmful actions. While a defensive development rather than a threat disclosure, it signals growing practitioner awareness of agentic AI risk and offers a concrete mitigation pattern for teams deploying coding agents like Claude Code, Codex, or Cursor.

Steganography in LLMs Enables Covert Data Exfiltration

ATLAS OWASP MEDIUM ▲ 6.5 Schneier on Security May 12, 2026

Research highlighted by Bruce Schneier confirms that LLMs are highly effective at embedding hidden messages within seemingly normal text, a technique known as text-in-text steganography. This capability raises significant concerns for covert communications, data exfiltration, and the evasion of AI content moderation systems. Even small models with ~4 billion parameters demonstrate robust encoding and decoding of obfuscated language, lowering the barrier for adversarial misuse.

Claude Chrome Extension Prompt Injection Enables Agent Takeover

ATLAS OWASP HIGH ▲ 9.1 SecurityWeek May 09, 2026

A vulnerability dubbed ClaudeBleed in Anthropic's Claude Chrome extension allows any browser extension to inject arbitrary prompts into the Claude AI agent by exploiting lax permission checks and improper trust validation. Attackers can bypass user confirmation protections via DOM manipulation and repeated message forging, enabling full agent takeover for information theft or unauthorized actions. The flaw effectively breaks Chrome's extension security model and exposes users running Claude's agentic capabilities to third-party extension compromise.

Researcher

Stay ahead of the threat.