Researchers have introduced EndoUT, an adversarial framework that identifies universal prompt triggers capable of significantly degrading LLM utility. By analyzing how a model's internal response-onset preferences are disrupted, attackers can create fixed strings that force performance drops across various benign tasks without needing target labels or feedback.
Google evaluated its Agent Anomaly Detection feature using five third-party benchmarks including AgentDojo and AgentHarm. The tool successfully flagged 87% of harmful instructions that agents actually executed, though it showed lower effectiveness against poisoned long-term memory.
IBM has introduced LLMON (LLM Object Notation), an XML-like markup language that uses reserved special tokens to explicitly label components like instructions, data, and constraints. Two experimental models, Granite 4.1 3B and 8B LLMON-preview, are now available, demonstrating near-zero success rates for indirect prompt injection attacks.
Amazon has released a defense-in-depth architecture to prevent memory poisoning—a form of indirect prompt injection—in AI agents that use persistent memory. The solution uses a Lambda-based interceptor on the AgentCore Gateway to validate writes using semantic drift, regex patterns, namespace boundaries, and rate limiting.
Recent research demonstrates that current security monitors often fail by confusing injection exposure with an agent's actual commitment to perform an unsafe action. The author proposes a double-stratified evaluation method that separates exposure from behavior, significantly improving the ability to predict unsafe tool calls up to two steps in advance.
A new study questions the common belief that whether an AI agent falls for a prompt injection is settled the moment it reads the malicious text. The researchers built a branching-rollout technique on AgentDojo benchmarks: they replayed an agent's run up to the point it encounters an injected instruction, then let the rest of the episode play out many different ways to measure how much of the final outcome was already locked in versus still open. For a Qwen3 agent running without extended reasoning, 91% of the outcome was already determined at that moment; switching on reasoning ('thinking') cut that figure to 62%, showing the agent's compliance decision often keeps forming well after it first sees the attack. Enabling reasoning also cut Qwen3's susceptibility to successful attacks, but made two other tested models more vulnerable, and benchmark scores for reasoning agents flipped on a rerun as often as 35% of the time.
Microsoft announced an expansion of its Project Perception security framework to address risks from autonomous AI agents, framing them as a new category of 'insider' that needs distinct identity, permission, and risk controls. The update introduces Dynamic Permissions (least-privilege, task-scoped, time-bound access), Dynamic Agent Risk (continuous risk scoring based on behavior drift and context), and Dynamic Conditional Access (graduated responses from narrowing access to full runtime isolation). Microsoft is also extending discovery and identity support (Microsoft Entra Agent ID) to 60 types of local agents, including Claude Code, OpenAI Codex, and GitHub Copilot, and introducing 'Guardian Agents' that monitor and respond to risky agent behavior at machine speed.
Cisco Talos' CAIRN research classifies a new malware archetype, 'A3: AI-Analysis Evasion,' where malicious code embeds natural-language text designed to manipulate language models used in automated malware triage. Talos traced the technique across four malware families (FRUITSHELL, PLOTSAFE, HOLLOWCLAD, MANTLEMAZE) and 84 samples from January 2025 to July 2026, showing it spreading from a simple copy-pasted comment to templated, multi-model-targeted 'template spraying' and fake intimidation messages. Controlled testing with a panel of five local LLMs found the simplest direct-instruction technique shifted model verdicts toward 'benign' in roughly 35% of runs, while more elaborate tricks had little effect or backfired by increasing suspicion.
Google Research released a report from its Contextual Agent Privacy and Security (CAPS) workshop, co-authored with over 50 academic and industry researchers. The report argues that AI agents need to reason about context-specific social norms, not just static permissions, to act appropriately when handling personal data and taking actions on a user's behalf. It proposes a multi-layered approach spanning system-level sandboxing, model-level reasoning, user-centric controls, multi-agent guardrails, and ecosystem governance, grounded in the theory of Contextual Integrity.