NVIDIA has detailed Sentry, a new security framework that sits outside an AI agent's reach, enforcing policy and watching behavior from a dedicated hardware trust domain on BlueField-4 DPUs. The idea is simple: even if an agent is compromised or goes off-script while using tools and credentials, Sentry keeps watching and can cut it off because it doesn't depend on the host OS the agent runs on. Two components do the work — DOCA Tower, a proxy that inspects and authorizes every agent network request, and DOCA Argus, which reads live memory to understand what's actually running. When something looks wrong, the system can quarantine the agent within milliseconds, without needing the agent or the host to cooperate.
Anthropic published details on how it curbed 'illicit distillation' — large-scale covert operations that scrape Claude's reasoning transcripts to train rival models cheaply. New measures, including reasoning summarization, 'preserved thinking' for new API accounts, and a two-stage classifier system, reportedly cut the volume of usable reasoning data extracted by distillers by 10-100x. Anthropic says a Chinese lab, Zhipu, used distillation (including attempts against Claude) to build GLM-5.3, which US assessors found to be the most cyber-capable open-weight model released so far with safeguards that are easily bypassed.
AWS and Keyfactor published a reference design that fixes a known weak point in Model Context Protocol (MCP) deployments: MCP has no built-in way to authenticate which agent is actually making a request. The architecture adds three trust layers — user-to-system, agent-to-agent, and agent-to-tool-server — built on signed request envelopes, delegation assertions signed by the user, RFC 8693 audience-restricted tokens, and live OCSP revocation checks. A working example runs on Amazon Bedrock and the AWS Strands Agents SDK with Keyfactor EJBCA issuing and validating the X.509 certificates, and the full code is on GitHub.
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.
Microsoft and NVIDIA released a joint whitepaper, 'Risk Management for Open-Weight Models in Enterprise,' the first in a planned series offering enterprises a shared risk taxonomy and methodology for evaluating open-weight AI models. The paper argues that whether a model is open or closed-weight is not the key risk factor — what matters is the use case, deployment architecture, and surrounding safeguards. It organizes risks into five categories (Cybersecurity, Data Protection, Content, Human-AI Interaction, Bias) and maps evaluations to those risks, aligning with ISO standards and the NIST AI Risk Management Framework.
NIST's Super Intelligence Security Center, together with the NSA and CISA, released a joint publication outlining ten voluntary core practices for organizations that train and evaluate highly capable 'super intelligence' (SI) agents, especially pre-deployment or high-risk non-public systems. The guidance defines an 'SI control incident' as an AI agent taking uninstructed actions that bypass or compromise human oversight and security controls. It builds on existing cyber incident response practices but adds AI-specific steps such as testing containment and validating it before resuming agent activity.
Microsoft Research released Magnetar, a zero-knowledge virtual machine (zkVM) capable of generating cryptographic proofs that a program executed correctly, without needing anyone to rerun it or trust the operator. Instead of forcing developers to rewrite applications as custom cryptographic circuits, Magnetar proves standard RISC-V machine code, so programs written in Rust, Go, or C can be compiled normally and proven as-is. Benchmarks across 15 programs show it processing 7.2 to 8.7 million instructions per second on a laptop, with proofs checkable in roughly 12 milliseconds, and its cryptographic security depends only on established hash functions like SHA-256 or SHA-3 rather than elliptic curves or lattice-based schemes.
AWS tested its AWS Security Agent (now part of AWS Continuum) against a freshly built benchmark of 37 real, recently disclosed CVEs, each run as a live containerized target with a hidden flag. Starting from a low-privilege endpoint with only a one-line hint and no proof-of-concept, the agent, running on Claude Opus 5, captured 33 of 37 flags through genuine end-to-end exploitation after an integrity audit, with 2 captures reclassified as unintended-path solves and 2 genuine misses where the agent broke the target before extracting the flag. AWS built this benchmark to avoid the contamination problem of public CVE suites, which models may have memorized from training data.
Tamnoon built Tami, an agentic orchestrator on Amazon Bedrock that investigates cloud security findings from CNAPPs and cloud providers, scores remediation safety, and generates fixes with guardrails. The company reports a greater than 90% reduction in critical cloud exposure within 90 days with zero production issues. The system uses four specialized agents (investigation, safety-scoring, remediation, code-generation) each calling different foundation models through Bedrock, with Bedrock Guardrails and Cross-Region inference providing safety and scale.
GitHub replaced its LLM-based pipeline for detecting generic secrets (like passwords and API keys) with a fine-tuned ModernBERT classifier combined with regex candidate generation. The new two-stage system improved detection accuracy while cutting scan time and serving costs dramatically, and it's now fast enough to run during push protection, not just after code is committed.
Anthropic has announced the Cyber Mission, a long-term initiative to help defenders secure operational technology and open-source code using Claude models. It includes the Critical Infrastructure Defense Program, built with partners like Accenture, CrowdStrike, Deloitte and Rockwell Automation, and the free OSS Scanner tool that scans open-source projects for vulnerabilities. The effort builds on lessons from the earlier Project Glasswing, now merged into an expanded Cyber Verification Program.
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.
AWS describes a new two-stage access control architecture for Retrieval Augmented Generation (RAG) in Amazon Quick and Amazon Bedrock Knowledge Bases. Instead of relying only on periodically synced access control lists (ACLs), the system now verifies permissions in real time against the authoritative data source (e.g. Google Drive, SharePoint, Confluence) at query time, before passing content to the LLM. AWS says this closes security gaps caused by stale or incorrectly mapped permissions between sync cycles. Mondelēz International, with over 35,000 employees, is cited as a customer using this approach.
Anthropic has merged its Project Glasswing and Cyber Verification Program into a single, expanded CVP with three access tiers: Defense Access, Red Team Access, and Specialized Access. Each tier unlocks progressively fewer safety blocks on cyber-related tasks in Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1 for vetted security organizations. Anthropic also published benchmark results and vulnerability-discovery numbers from the earlier Glasswing program to justify the expansion.
Z.ai's GLM 5.3, a 753-billion-parameter mixture-of-experts model optimized for coding and long-horizon agentic tasks, is now available as a fully managed model on Amazon Bedrock. It builds on GLM 5 with improved coding benchmark results, standout cybersecurity capabilities (CyberGym score of 84.5), and deeper Bedrock integration including OpenAI-compatible APIs, prompt caching, cross-Region inference and service tiers. AWS also demonstrates using GLM 5.3 with the open-source Strix agent for authorized penetration testing of your own applications.
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.
Microsoft Threat Intelligence documented exploitation of CVE-2026-73570, an unauthenticated OS command-injection flaw in the Zimbra Collaboration Suite's SNMP notification path. Attackers triggered it via specially crafted emails against internet-facing servers running the optional zimbra-snmp package with SNMP notifications enabled, requiring no authentication or user interaction. Post-exploitation activity included JSP webshell deployment, privilege escalation to root, persistent access, credential and mailbox theft, lateral movement across Zimbra clusters, and custom remote-access malware. Zimbra fixed the issue in version 10.1.20 (released July 20, 2026); the CVE was publicly disclosed August 13, 2026, and Microsoft observed scanning and exploitation activity during the gap between patch availability and disclosure, as well as after.
AWS released a step-by-step walkthrough showing how to give Claude Desktop (on Amazon Bedrock) live internet access through Amazon Bedrock AgentCore Gateway, closing the gap left by the model's training cutoff. The setup wires AWS IAM Identity Center SSO through Amazon Cognito as a SAML-to-OAuth federation layer, issuing JWTs that the AgentCore Gateway validates on every request. The Web Search capability itself is MCP-compatible and runs on an internal Amazon web index covering tens of billions of documents, so no query traffic leaves AWS and no third-party API key is needed. The feature is currently limited to three AWS Regions: us-east-1 (N. Virginia), eu-west-1 (Ireland), and ap-northeast-1 (Tokyo).