Every article scored, classified, and mapped to MITRE ATLAS and OWASP LLM Top 10 — so you always know what matters and why.
AI security intelligence analysis for 2026-W35 — MITRE ATLAS technique trends, OWASP LLM risk distribution, threat actor activity, and enterprise readiness assessment based on 25 articles.
Read full analysis →
Anthropic is expanding its Claude Cowork agentic desktop feature to mobile, enabling users to remotely initiate, monitor, and steer long-running AI tasks on their PC from a smartphone — with background task execution persisting even after the mobile app is closed. This cross-device architecture introduces a new attack surface: a mobile application acting as a command-and-control interface for an agent with local filesystem access, expanding the blast radius of device compromise, session hijacking, and prompt injection attacks. Defenders must now account for a persistent, background-running agentic process on employee endpoints that can be triggered or manipulated via a separate, potentially less-secured mobile channel.
A malware developer has embedded nuclear and biological weapons-related text inside JavaScript comment blocks within spyware payloads, specifically to trigger refusal behaviour or context confusion in LLM-powered security analysis pipelines. The technique exploits the architectural gap between how interpreters (which skip comments) and language models (which ingest the full file as input) process the same file. While ineffective against traditional static analysis tooling, the tactic represents a practical adversarial countermeasure targeting AI-first triage workflows and analyst copilots.
OpenAI has unveiled 'Jalapeño', its first custom-built AI inference processor co-designed with Broadcom, optimised for running large language models at reduced cost and power consumption. The move deepens OpenAI's vertical integration across the full AI stack — from chip silicon through to end-user products — introducing new hardware supply chain dependencies and firmware-level attack surfaces that defenders must now account for. Security teams should treat purpose-built AI silicon as a new tier of the ML supply chain, with unique risks around hardware backdoors, firmware integrity, and reduced hardware diversity.
Google DeepMind researchers have released a structured taxonomy categorising adversarial attacks against autonomous AI agents into six classes — content injection, semantic manipulation, cognitive state poisoning, behavioural control, systemic, and human-in-the-loop traps — formalising an emerging threat model for agentic AI systems. For defenders, this framework codifies attack paths that exploit the agent's inability to distinguish trusted instructions from attacker-controlled data ingested from web pages, emails, documents, and tool outputs. NIST evaluation data cited in the research shows malicious instruction injection succeeded in 57% of tested agent hijacking scenarios on average, underscoring that these are active, high-yield attack vectors rather than theoretical concerns.
Agentic AI systems deployed in security operations and enterprise workflows are increasingly executing autonomous decisions at machine speed, using LLM-derived confidence regardless of context accuracy. The core security risk is that incomplete, poisoned, or manipulated context fed to these agents produces confidently wrong actions executed without human review. Defenders face a compounded threat: adversaries can now target the context layer—asset inventories, threat feeds, exposure data—to induce systematic misconfiguration or inaction at scale.
MoEngage has acquired Aampe to deploy individualized AI agents for every customer, enabling autonomous decisions on messaging targeting, timing, and content at enterprise scale across 1,350+ brands globally. This architecture introduces a large, distributed fleet of autonomous agents operating on sensitive behavioral and PII data, dramatically expanding the blast radius of any single compromise. Security teams at enterprises adopting this platform must now reason about agent-level trust boundaries, data inference risks, and the amplification potential of adversarial manipulation across millions of simultaneous decision-making agents.
Dragos has launched EmberAI, an AI module embedded within its OT security platform that allows analysts to query threat intelligence, asset data, and network activity in plain language, grounded in a decade of proprietary OT-specific data. The system introduces new attack surface considerations because it aggregates highly sensitive OT network telemetry, vulnerability data, and adversary intelligence into a single AI-queryable layer — making the platform itself a high-value target. Defenders must weigh the risks of prompt injection, over-reliance on AI-generated recommendations in safety-critical environments, and the intelligence value this consolidated dataset represents to nation-state adversaries.
Mistral OCR 4 is a production-grade document intelligence model delivering bounding boxes, block classification, inline confidence scores, and 170-language OCR optimised for enterprise RAG and search ingestion pipelines. For defenders, the model's role as a trusted ingestion component in downstream retrieval pipelines creates a high-value attack surface: adversarially crafted documents can now influence RAG context, citations, and automated redaction decisions at scale. The self-hosted single-container deployment option further expands the supply chain and misconfiguration risk surface for organisations running document intelligence internally.
A campaign dubbed 'Cordyceps' is exploiting weaknesses in CI/CD workflows to inject malicious pull requests into high-profile open-source projects, including Google's AI Agent Development Kit and Microsoft's Azure Sentinel. The attack surface spans multiple trusted ecosystems, meaning poisoned code could propagate into AI tooling, cloud infrastructure, and widely-used developer utilities before detection. The breadth of targets — including Python's Black formatter — signals a supply chain strategy designed to maximise downstream blast radius.
Anthropic's Mythos AI model identified vulnerabilities in classified US government computer systems within hours during a government-sanctioned testing exercise under Project Glasswing. A senior US official confirmed the findings to the Associated Press, corroborating statements made by Sen. Mark Warner that the model 'broke into almost all of our classified systems.' The incident marks a landmark demonstration of AI-enabled offensive cyber capability at the highest sensitivity levels of government infrastructure.
Security firm AIR demonstrated that a malicious AI agent skill, disguised as a Google Stitch landing-page builder, passed every major skill scanner including Cisco's, NVIDIA's, and skills.sh integrations, reaching approximately 26,000 agents before its payload was activated. The attack exploits a structural gap: scanners evaluate a static package at submission time, while the external URL the skill instructs the agent to fetch can be silently swapped post-install to deliver arbitrary instructions. Defenders relying on marketplace reputation signals, GitHub star counts, or one-time scanner verdicts to gatekeep agent skills have no meaningful protection against this class of supply-chain attack.
Attackers are bypassing AI-layer defences entirely by exploiting unpatched legacy infrastructure — misconfigured Active Directory, stale credentials, and over-privileged IAM roles — to hijack the resources AI agents depend on. Research cited in the article shows 70% of organisations grant AI systems more access than a human in the same role, driving a 76% incident rate among over-privileged deployments. The article argues that securing AI agents requires closing the underlying infrastructure exposure gap, not just hardening the model layer.
New research from Ye, Cui, and Hadfield-Menell demonstrates that LLMs prioritise the stylistic format of text over its structural role tags, enabling attackers to craft injected content that mimics internal reasoning blocks and bypasses safety guardrails. The study found attack success rates of 61% when injected text stylistically matched model-internal formats, dropping to just 10% after 'destyling'. The authors conclude that without genuine role perception in models, prompt injection defences will remain fundamentally reactive.
OpenAI has partnered with Trail of Bits to launch 'Patch the Planet,' an initiative using AI-assisted tooling (including Codex Security) to help open-source maintainers find and patch vulnerabilities at scale. While the defensive intent is clear, the program introduces new attack surface considerations: AI-generated patches applied to widely-used open-source projects create a high-value supply chain target, and the triage/remediation pipeline itself could be manipulated to introduce subtle flaws. Defenders should monitor open-source dependencies that receive AI-assisted patches and assess the integrity guarantees of the remediation workflow.
A three-flaw vulnerability chain dubbed AutoJack in Microsoft's AutoGen Studio allowed attackers to execute arbitrary commands on a developer's host system by manipulating a browsing AI agent into connecting to a malicious webpage. The attack exploited missing authentication on MCP WebSocket routes combined with unsanitised base64-encoded parameters to launch arbitrary processes. Microsoft confirmed the flaw was patched before any PyPI release, limiting exposure to developers building directly from the main GitHub branch.
◉ AI THREAT BRIEFING
Twice-weekly digest of critical AI security developments — every story mapped to MITRE ATLAS and OWASP LLM Top 10. Free.
No spam. Unsubscribe anytime.