All news in category “AI and Security Pulse”

🔍 While many organizations intentionally deploy AI to improve productivity, unsanctioned AI is proliferating faster — employees install tools or vendors embed assistants into existing apps. The article defines four AI categories and maps specific detection techniques to each, covering DNS, web gateways/NGFW, EPP/EDR, application and browser controls, and SSPM/identity governance. It flags OAuth consent as a high-risk channel and summarizes admin steps for Microsoft Entra, Google Admin, Salesforce, and ServiceNow to block or restrict app access.

🔍 ENISA chief Hans de Vries told ESET World that AI-powered vulnerability scanners mean firms can no longer claim ignorance of software bugs. He warned that the Cyber Resilience Act and emerging AI tools require security by design and that failure to use AI coherently risks exploitation and litigation. The NCSC also expects AI to expose poorly coded systems while vendors adopt AI to remove flaws.

🔍 Across organizations, employees run three to five AI tools daily—many unapproved and often connected to corporate data via OAuth, browser extensions, or newly added vendor features—creating a widening "shadow AI" gap that evades traditional network controls. The article outlines five practical steps security teams can apply: build an inventory, write usable policies, create a fast approval lane, implement browser-native monitoring, and deliver just-in-time coaching. Together these measures aim to preserve productivity while restoring visibility, reducing data exposure, and aligning employee workflows with security requirements.

🔍 New benchmarks from the UK’s AI Security Institute (AISI) show leading AI models rapidly improving at multi-stage penetration testing, with the difficulty of tasks solvable by models doubling every 4.7 months as of early 2026. The tests measure the longest task an AI can complete with 80% success relative to human work-hours, emphasizing autonomous chaining of steps rather than raw speed. While there are caveats — token limits and inconsistent model performance — the findings highlight growing offensive and defensive implications for enterprise security.

🔍 Researchers from Xidian University describe a new image-based prompt injection called CrossMPI that uses near-imperceptible pixel perturbations to alter how large vision-language models interpret both visual and textual inputs. The technique targets intermediate multimodal fusion layers rather than final outputs, misleading LVLMs without modifying text prompts. Tests show strong black-box transferability and high success rates across several open-source models, while common defenses reduce but do not fully eliminate the threat.

🔒 The UK’s National Cyber Security Centre (NCSC) has published new guidance for organisations considering the adoption of agentic AI, summarising a wider report produced with Five Eyes partners. It flags the heightened risk from agent autonomy and complexity, including excessive access, unpredictable behaviour and actions that can outpace human review. The NCSC advises incremental deployment with tightly bounded pilots, clear ownership, ongoing monitoring and meaningful human oversight, and points organisations to industry best practice such as ETSI EN 304 223.

🛡️ The rapid rise of AI-assisted and vibe coding is accelerating secrets sprawl, with developers and AI agents increasingly introducing credentials, tokens, and private data into code and collaboration tools. Security researchers from Wiz and independent analysts found a Jan. 28, 2026 Moltbook backend misconfiguration on Supabase that exposed 1.5 million API authentication tokens, tens of thousands of emails, and private messages. Organizations report that detection is outpacing remediation: many teams can find leaks but lack governance and processes to revoke, rotate, and purge secrets at scale. Experts urge treating the issue as identity governance, embedding security into the SDLC, and enforcing short-lived credentials and automated rotation.

🔍 Cloudflare tested security-focused LLMs on its infrastructure and reports detailed findings from using Anthropic’s Mythos Preview as part of Project Glasswing. The model stood out for exploit chain construction and automated proof generation, producing runnable PoCs and iterating on failures. Its emergent guardrails proved inconsistent across runs and prompts, so Cloudflare built a tailored harness and additional safeguards to scale safely. The team also observed higher-quality, actionable findings compared with earlier frontier models, but noted increased noise from memory-unsafe languages and model bias.

🔒 The article argues that security must move beyond detection and focus on compressing the OODA loop—observe, orient, decide and act—so defenses can outrun attackers. It notes that detection improvements have reached diminishing returns while investigation and remediation remain time-bound bottlenecks. By embedding contextual investigation into systems and deploying agent-based remediation, teams can make faster, more consistent decisions. As AI-driven interactions accelerate threat timelines, continuous validation and automated response become essential.

🆕 AWS added FLUX.2-klein-base-4B and Qwen3-Embedding-0.6B to Amazon SageMaker JumpStart. FLUX.2 targets real-time image generation and multi-reference editing in a compact architecture that can run on consumer GPUs with about 13GB VRAM. Qwen3-Embedding delivers instruction-aware, multilingual text embeddings across 100+ languages for retrieval, RAG, and semantic search. Customers can deploy these models via SageMaker Studio or the SageMaker Python SDK.

🛡️ Microsoft Security explains how rising agentic autonomy reorients security from models to how agents are assembled, constrained, and governed inside applications. The post identifies amplified risks—agent hijacking, intent breaking, data leakage, supply chain compromise—and shows why the application layer is decisive because builders fully control permissions, tool access, and failure handling. It recommends concrete design patterns: agents as microservices, least permissions, deterministic human-in-the-loop, and distinct agent identity to limit blast radius and preserve auditability.

⚠️ AI hallucinations—confident but incorrect outputs—are increasingly driving risky decisions in critical infrastructure and cybersecurity operations, exploiting human trust in authoritative-sounding responses. A 2025 AA-Omniscience benchmark of 40 models found most systems were more likely to offer a confident wrong answer on difficult questions, underscoring that AI outputs must be treated as potential vulnerabilities until vetted. Effective controls include enforced human review before sensitive actions, treating training data as a security asset, strict least-privilege for AI systems, and prompt-engineering training to reduce ambiguous inputs.

🔍 Anthropic’s announcement that Claude Mythos Preview will not be released publicly underscores both genuine capability and strategic constraint. Independent testing and reproductions suggest similar performance from OpenAI’s GPT-5.5 and smaller community models, while Mythos’ cost and corporate incentives shape access. These generative systems dramatically improve automated vulnerability discovery, empowering both attackers and defenders. Mozilla’s use found 271 flaws, but many devices remain unpatchable, so organizations must adapt quickly.

🔒 Semperis finds that 93% of global organizations use or plan to use AI agents for security tasks such as password resets and VPN access, while 92% report AI on endpoints with SSH and encryption key access. The survey of 1,100 organizations warns of over‑permissioned and abandoned 'zombie' non‑human identities that increase hijack risk. Semperis recommends treating agents as NHIs, enforcing least‑privilege, and improving observability and recovery readiness.

🔒 Palo Alto Networks reports ongoing testing of frontier AI models, including Anthropic and OpenAI, finding they rapidly surface code vulnerabilities and potential exploit paths. In the May 'Patch Wednesday' advisories the majority of findings originated from these AI scans, prompting broad rescanning and remediation. The company warns of a narrow three-to-five-month window before AI-driven exploits spread and offers Unit 42 services to help organizations respond.

🛡️ Microsoft introduced MDASH (multi-model agentic scanning harness), a model-agnostic AI system in limited private preview designed to discover, validate, and prove exploitable defects in large codebases. The system orchestrates more than 100 specialized agents across frontier and distilled models in a structured pipeline that builds threat models, runs auditor and debater stages, groups equivalent findings, and proves vulnerabilities. Microsoft reports MDASH uncovered 16 issues fixed in this month’s Patch Tuesday, including two critical Windows networking and authentication flaws.

🔒 Anthropic's Mythos Preview, shared last month with a limited set of security partners, has demonstrated the ability to autonomously find zero-day vulnerabilities across major operating systems and browsers. Anthropic paired the release with Project Glasswing and $100 million in usage credits to help defenders, but reports of unauthorized access and denied requests from Chinese entities have already emerged. The development challenges the assumption of a durable US lead and has injected cybersecurity into high-level US–China summit talks, prompting urgent questions about access, regulation, and international cooperation.

🛡️ In a post-Mythos environment, AI-driven attacks can weaponize vulnerabilities within hours or minutes, outpacing traditional defensive cycles. Picus Security argues defenders must pair continuous Breach and Attack Simulation (BAS) with autonomous pentesting to validate controls and reveal genuine attack paths. Operational friction — the "spaghetti handoff" between tools and teams —, not tooling alone, is the main cause of delayed response, so validation must be automated end-to-end.

🔍 The UK’s AI Security Institute evaluated GPT-5.5’s ability to identify software security vulnerabilities and concluded it performs comparably to Claude Mythos, based on a series of red-team style tests and benchmark prompts. The assessment highlights that GPT-5.5 is generally available from OpenAI, making high-quality automated vulnerability detection more accessible to organizations and researchers. The Institute also analyzed a smaller, cheaper model which, when given additional prompting scaffolding and careful supervision, delivered similar detection performance. Overall, the study suggests parity among leading LLMs for initial vulnerability discovery, with differences largely hinging on prompt engineering and deployment context.

🔒 Microsoft researchers describe an AI-driven pipeline that translates attacker TTPs into realistic, structured security logs to accelerate detection engineering. The approach uses prompt engineering, collaborative agentic refinement, and data augmentation to generate semantically accurate telemetry (command lines, process ancestry, fields) without exposing sensitive customer data. Evaluation across multiple datasets shows agentic workflows and reasoning models notably improve recall and fidelity compared to prompt-only methods.