All news in category “AI and Security Pulse”

🤖 A new laboratory experiment examines how humans respond when pitted against LLMs in a multi-player p-beauty contest versus other humans. Using a within-subject, monetarily-incentivised design, the study finds participants choose significantly lower numbers when playing against LLMs, with a marked increase in selections of the zero Nash-equilibrium. The effect concentrates among participants with strong strategic-reasoning ability, who report perceived AI reasoning and an unexpected expectation of cooperation as motivating factors.

🔐 CrowdStrike has been selected for OpenAI's Trusted Access for Cyber (TAC) program and will integrate the frontier model GPT-5.4-Cyber into its platform. Its multi-model AgentWorks framework enables defenders to choose the best model for each task while applying enterprise-grade governance and real-world threat intelligence. Falcon sensors provide runtime visibility across endpoints, governing AI agents where they execute and helping organizations meet emerging regulatory requirements such as the EU AI Act.

🔒 Smashing Security episode 463 examines two incidents that expose operational and AI security weaknesses: a claimed intrusion into Venice’s flood‑defence pump controls and an accidental full‑source disclosure by Anthropic. Hosts Graham Cluley and Tanya Janca discuss the physical risks of compromised legacy OT systems, how packaging/CI misconfigurations can leak high‑value IP and attack surface, and the governance challenges of powerful internal tools like Mythos. They recommend stronger CI/CD defaults, strict access controls for model assets, and reliable out‑of‑band incident communications.

🧑‍🦯 Be My Eyes and its Be My AI feature can help visually impaired users identify on-screen content and even flag phishing attempts, but they are not infallible. In tests, the AI identified fake login pages and suspicious emails, yet risks such as hallucinations and prompt-injection remain. Treat AI output as a first-pass check, avoid sharing confidential details with unknown volunteers, install trusted security software and use a password manager, and prefer apps that process sensitive documents locally when possible.

🔍 AI changes the assumptions behind incident response: outputs are non-deterministic, harmful content can be produced at machine speed, and root causes often emerge from interactions among training data, fine-tuning, retrieval, and user context rather than a single code defect. The familiar principles of explicit ownership, containment before investigation, psychologically safe escalation, and clear communication still apply, but teams must expand taxonomies and severity frameworks to capture AI-specific harms. Closing gaps in observability, reconciling privacy defaults with forensic needs, and adopting staged remediation—stop the bleed, fan out and strengthen, and fix at the source—are critical, as is protecting responder wellbeing during prolonged incidents.

🔒 OpenAI has released GPT-5.4-Cyber, a variant of GPT-5.4 fine-tuned to assist cybersecurity tasks and defensive workflows. The company expanded its Trusted Access for Cyber (TAC) program with tiered verification so vetted defenders can access models with lowered refusal boundaries for legitimate security work. Access is currently limited to vetted vendors, organisations and researchers while OpenAI carefully studies benefits and risks. The release is positioned to embed advanced coding and agentic capabilities into secure development practices.

🔒 Organizations must treat AI as part of their core architecture rather than a separate stack. Effective protection extends existing controls — identity, policy enforcement, observability, and data governance — across AI interfaces, private LLMs, and agentic systems. Security requires coordinated runtime enforcement at firewalls, API gateways, and SIEM with zero-trust principles. Fortinet positions converged platforms as the way to embed AI guardrails into the foundational operating model.

🤖 Agent Lee is an in-dashboard AI assistant that understands your Cloudflare account and can troubleshoot, visualize, and modify resources on demand. Built on Cloudflare primitives like Workers, Durable Objects, and the Agents SDK, it uses a TypeScript Codemode to generate server-side calls that run through a sandboxed MCP proxy. Write operations require explicit user approval via an elicitation gate, and the beta — serving ~18,000 daily users and ~250k tool calls per day — is available now.

⚠️ Security researchers at Capsule Security disclosed prompt-injection vulnerabilities in Microsoft Copilot Studio and Salesforce Agentforce that let attackers embed malicious instructions in public form fields. Crafted inputs submitted via SharePoint or lead forms can override agent instructions and trigger data exfiltration to attacker-controlled endpoints. Microsoft patched the SharePoint-related issue (CVE-2026-21520) with a 7.5 CVSS score; Salesforce acknowledged the problem but described the vector as configuration-specific. Researchers warn that treating external inputs as trusted undermines autonomous agent security and urge input validation, least-privilege, and stricter outbound controls.

🔒 AI adoption is now a boardroom expectation, and Pentera’s AI Security and Exposure Report 2026 reports that every CISO surveyed already uses AI across their organizations. The piece argues that fully agentic systems, while powerful and adaptive, introduce probabilistic variability that undermines repeatable, measurable security validation. A hybrid approach—deterministic orchestration for consistent attack chains combined with AI for adaptive payloads and environmental interpretation—provides guardrails while preserving realism. This anchoring enables reliable retesting and continuous exposure validation without sacrificing contextual intelligence.

🔒 At VulnCon26 in April, Lindsey Cerkovnik of CISA urged that AI firms like OpenAI and Anthropic be more directly represented in the CVE program to help manage a surge in reported vulnerabilities. She warned that new AI tools both accelerate discovery of valid flaws and generate lower-value noise, putting pressure on disclosure workflows. Recent vendor developments — Anthropic’s Mythos Preview and OpenAI’s GPT-5.4-Cyber — illustrate how automated research is already changing the threat landscape. Cerkovnik said CVE funding is secure and the program remains a CISA priority.

🔎 Deepfake technology is now widely accessible and sufficiently realistic to fool employees, executives and automated heuristics. A 2025 Gartner survey found nearly half of cybersecurity leaders encountered audio or video deepfakes in the prior year, and real-world incidents show attackers using synthetic media for both financial fraud and reputational sabotage. Organizations must combine rapid forensic verification, coordinated legal action and clear communications while pursuing long-term authentication and watermarking standards to restore trust.

🛡️ OpenAI has unveiled GPT-5.4-Cyber, a variant of its flagship GPT‑5.4 tuned for defensive cybersecurity use cases, and expanded its Trusted Access for Cyber (TAC) program to include thousands of authenticated individual defenders and hundreds of security teams. The company says the model is intended to help teams find, validate, and fix vulnerabilities faster while it iteratively strengthens safeguards to reduce dual‑use risks and resist jailbreaks and adversarial prompt injection. OpenAI highlighted its Codex Security agent, which it credits with contributing to the remediation of over 3,000 critical and high vulnerabilities, and framed the release as part of a broader shift toward continuous, developer‑integrated security feedback.

🔒 Curity announced Access Intelligence, an extension to its Identity Server IAM platform designed to secure rapidly proliferating autonomous AI agents. Rather than rely on static, pre-granted permissions, the company uses Token Intelligence to embed an agent's declared purpose and intent in OAuth tokens and issues short-lived, action-specific tokens at runtime. The system can require human approval for high-risk tasks, is deployed as a self-hosted microservice, and centralizes token validation to isolate unregistered or shadow agents.

🔒 This post explains how AI agents and coding assistants access AWS resources via the Model Context Protocol (MCP) and why deterministic IAM controls are required. It outlines three security principles—assume all granted permissions could be used, enforce role governance, and differentiate AI-driven from human-initiated actions—and maps them to deployment patterns. It contrasts AWS-managed MCP servers (which inject context keys) with self-managed servers (which require session tags), and provides practical IAM policy examples, monitoring guidance, and operational controls.

🛡️ AI is reshaping cyber resilience, accelerating both innovation and adversary capabilities. Organizations must move beyond static perimeter defenses to a model of continuous cyber resilience, emphasizing always-on monitoring, automation, and rapid recovery. Platform consolidation, human-centric operations, and regulatory reporting will define the next 3–5 years.

🔒 European regulators have been largely frozen out of early access to Anthropic's new Mythos model, Politico reports. Anthropic's Project Glasswing has initially restricted testing to select U.S. technology firms — notably Apple, Microsoft and Amazon — so partners can evaluate and mitigate security risks. The UK’s AI Security Institute has been permitted to test Mythos and acted on findings, while Germany has opened dialogue but not gained access, prompting concerns about private-sector control over a potent security-focused AI.

🔐 Cloudflare outlines how it centralized and secured company-wide use of the Model Context Protocol (MCP) by combining controls from Cloudflare One and its developer platform. The post explains why locally hosted MCP servers posed supply‑chain and administration risks and how Cloudflare moved to governed, remote MCP servers with Access-based authentication, audit logging, and CI/CD templates. It highlights MCP server portals, Code Mode to reduce token costs, and Gateway detection for shadow MCP.

🧠 A new paper, What hackers talk about when they talk about AI, analyzes more than 160 cybercrime forum conversations collected over seven months to show how offenders perceive and experiment with AI. The study finds growing curiosity about using both legitimate AI services and bespoke illicit models, alongside clear doubts about reliability, cost, and operational security. Authors use a diffusion-of-innovation framework to trace early-stage adoption and offer practical guidance for law enforcement and policymakers.

🔐 The UK’s AI Security Institute (AISI) evaluated Anthropic’s Claude Mythos Preview and found it can autonomously discover and exploit vulnerabilities in controlled tests when given network access. In a 32‑step simulated corporate attack the model completed the full sequence in 3 of 10 runs and averaged 22 of 32 steps, though performance varied. AISI stresses these cyber ranges are easier than real environments and recommended organisations strengthen basics — timely patching, robust access controls, secure configuration and comprehensive logging — while also exploring AI to bolster defensive capabilities.