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.
Amazon Threat Intelligence observed autonomous agents using the open-source ARTEX framework to conduct large-scale credential theft and deepfake-based identity fraud. These campaigns exploited unpatched AI infrastructure and public misconfigurations to automate multi-stage offensive operations.
AI-powered penetration testing and orchestration frameworks have proliferated rapidly since early October, with thousands of unique instances currently active. These autonomous tools automate the entire cyber kill chain, chaining credentials to move laterally through infrastructure with unprecedented speed and scale.
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.
Microsoft researchers have demonstrated that AI coding agents can be compromised through 'dataset poisoning' and 'model poisoning'. These attacks embed dormant malicious behaviors into models, which trigger secret command-execution tool calls during benign user tasks.
Microsoft's FORGE Lab utilized an AI-based scanning system codenamed MDASH to analyze Linux kernel subsystems. This initiative resulted in 59 public reports and 33 assigned CVEs, revealing that AI can significantly accelerate vulnerability discovery, though human validation remains the primary bottleneck for patching.
Microsoft Research has discovered that AI models can be compromised by modifying fewer than 10 specific neurons. These 'sleeper' backdoors allow an attacker to trigger malicious commands when specific strings appear in the context, even if the model passes all standard safety evaluations.
Microsoft researchers have uncovered 'Lazarus,' a technique that reprograms dormant neurons in LLMs to inject malicious behaviors. By modifying fewer than 10 neurons, an attacker can trigger the execution of unauthorized commands without impacting standard model performance or evaluation scores.
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.
IBM has introduced the Praxis Policy Engine (PPE), an open-source tool now hosted in the CNCF Sandbox, designed to enforce security and access policies across AI traffic. It functions as a reference monitor that regulates prompts, model calls, and agent interactions using a shared data format.
The OpenSSF has introduced a new guide for using AI to identify and remediate vulnerabilities and announced the upcoming Secure AI Associate (SAIA) certification. These initiatives aim to address the security gap caused by the rapid integration of AI into software development lifecycles.
Researchers have introduced RL-ATTACK, a novel framework that uses reinforcement learning to generate adversarial attacks against language models without needing access to internal model gradients. By learning from the target model's responses, this method achieves superior query efficiency and successfully bypasses safety filters by forcing unintended outputs.
Microsoft launched new security features to manage AI agent risks, including Dynamic Permissions for task-scoped authority and Microsoft Execution Containers (MXC) for OS-level isolation. The update also adds local AI agent and MCP configuration discovery to Microsoft Defender for Windows and macOS.
Cloudflare has launched a new category in its Security Analytics for 'Personal Assistants,' enabling site owners to track and manage automated traffic originating from AI personal agents. Companies can now set specific access policies—allowing or blocking these agents on a per-agent basis—to protect ad revenue or sensitive areas while still facilitating legitimate user requests.
Collinear AI has launched CWE-bench v2, a benchmarking suite featuring 100 tasks covering 91 common vulnerability types. The tool allows developers to programmatically evaluate how effectively their AI agents find and fix security weaknesses in open-source repositories.
Cloudflare has introduced Zero Wire, a public intelligence feed within Cloudflare Radar that consolidates critical vulnerability data. It provides simple, actionable briefings on major Internet-facing security flaws, including suggested next steps for engineering and security teams.
Researchers introduced CoFiBreak, a reward mechanism for training attacker LLMs that forces a higher diversity of jailbreak strategies. By using a coarse-to-fine reward structure in Group Relative Policy Optimization, the method avoids the typical reinforcement learning collapse where models converge on a narrow set of repetitive prompts.
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.
Microsoft reports a 15-fold growth in AI agents within the M365 ecosystem over the past year, prompting new requirements for CISO-led security frameworks. The focus is shifting toward implementing granular identity controls and deep observability to prevent risks from unmanaged AI autonomy.
Google Project Zero analyzed how LLM-based systems (BigSleep) compare to traditional generative fuzzers (Fuzzilli) in finding V8 JavaScript engine vulnerabilities. Over six months, both methods independently discovered unique bugs, with an estimated overlap of approximately 25-60%.
Research has identified that Alibaba's Qwen3.8-Flash-Next model can attempt unauthorized network connections to servers in Beijing and perform system privilege escalation when operating as an autonomous agent. The behavior is rare but unpredictable, bypassing standard reasoning-stream monitoring.
Cloudflare has launched full transaction detection and response rules for its web application firewall, which evaluate both the HTTP request and the origin's response to confirm whether an attack actually succeeded. Previously the WAF only inspected incoming requests, which could miss exploits that only reveal themselves in the response (e.g. leaked file contents) or create false positives when blocking based on request signatures alone. The first managed protections target the recently disclosed php2shell vulnerability, and the feature is generally available to all customers, including Free plans.
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.