All news in category “AI and Security Pulse”

🔒 OpenAI’s Lockdown Mode aims to reduce AI-enabled data exfiltration by disabling web browsing, image support, Deep Research, Agent Mode, network access from generated code, and file downloads while still permitting manually uploaded file analysis. Experts say the feature is a pragmatic but imperfect mitigation that still allows side-channel exfiltration, complicates governance across multiple AI vendors, and shifts responsibility between providers and enterprise security teams.

🔒 The White House Executive Order on advanced AI seeks practical public–private coordination to address AI-driven cyber risks while preserving innovation. It prioritizes voluntary model assessments, improved federal defenses, faster vulnerability discovery and remediation, and expanded cybersecurity talent. Successful implementation will hinge on operationalizing AI-assisted defense, translating insights into timely guidance and mitigations, and supporting resource-constrained critical infrastructure operators.

📰 Anthropic launched Project Glasswing in April to let companies use its Mythos model to discover and remediate software vulnerabilities. The project produced a status report claiming many findings, including some dangerous issues, yet most reported vulnerabilities appear unpatched. Anthropic’s reluctance to release detailed data and methodology — instead asking the public to "trust us" — raises questions about the accuracy and interpretation of the results.

🛡️ Ariel Fogel warned at Infosecurity Europe 2026 that prompt injection is an unresolved architectural issue threatening AI development. He explained that LLMs treat inputs as a single token stream, preventing reliable privilege separation between system prompts, user inputs and agent-retrieved content. With agents gaining tool access, successful injections can escalate from bad outputs to real-world actions, outpacing traditional governance and controls.

🧭 Microsoft has released an open-source fork of Windows Terminal named Intelligent Terminal, enabling AI assistance directly within the terminal without disrupting active sessions. The assistant can explain errors, draft commands, and propose fixes while remaining aware of current and past agent sessions. Users choose an AI agent (examples include GitHub Copilot, Claude, Codex, and Gemini) and can toggle Automatic error detection, Automatic error suggestion, and Session management. The terminal shows an AI pane beneath the shell for interactive planning, edits, and session resume features.

🔒 OpenAI has started rolling out a new Lockdown Mode for eligible ChatGPT personal accounts to reduce the risk of data exfiltration from prompt injection attacks. The optional security setting restricts capabilities that can connect to the web or external services, including live web browsing, image support, agent mode, deep research, Canvas networking, and file downloads. Lockdown Mode is available across Free, Go, Plus, Pro, and self-serve ChatGPT Business plans but cannot be used simultaneously with Developer Mode. OpenAI warns the feature reduces but does not eliminate exfiltration risk and also launched enhanced account session management to help detect and terminate unauthorized access.

🔍 Microsoft has expanded its Taxonomy of Failure Modes in Agentic AI Systems with seven newly identified ways agentic AI can be compromised. The update cites rapid adoption, maturation of the Model Context Protocol (MCP) ecosystem, proliferation of computer-use agents, and increased empirical evidence as drivers. New failure modes include supply chain compromise, goal hijacking, inter-agent trust escalation, visual attacks on CUAs, session context contamination, MCP/plugin abuse, and capability disclosure. Microsoft recommends inventorying agent supply chains, issuing cryptographic attestations, adding these modes to red-team exercises, and auditing human-in-the-loop controls.

🛡️ Lloyds Banking Group treats agentic AI as an engineering problem to be designed, constrained and tested at scale. At OWASP’s GenAI Security Summit, Lloyds’ security leads explained an “AI safe adoption” strategy spanning lifecycle governance, an internal agent marketplace, and multidisciplinary feature teams. Key challenges include agent identity, runtime observability and automated red‑teaming, while prioritizing low‑risk, high‑value use cases for customers.

🛡️ The Open Worldwide Application Security Project (OWASP) published a new agentic AI security maturity framework in the GenAI Security Project paper "State of Agentic AI Security and Governance" on June 3, and introduced it at Infosecurity Europe 2026 on June 4. The Enterprise Adoption Maturity Model maps deployments (from shadow AI to multi-agent systems) against governance maturity (from ad hoc to continuous oversight). It provides a decision tool to identify mismatches and prescribes either tailored controls for agentic systems or constrained agent permissions until governance catches up.

🔒 Ox Security urges that appsec be integrated directly into AI coding tools as agentic development accelerates code changes beyond traditional pipelines. Speaking at Infosecurity Europe, field CTO Boaz Barzel argued that security must become a continuous, contextual property of creation rather than a bolt-on stage. He outlined four agentic attack surfaces—input, tools, execution and output—and advocated autonomous security agents that pentest and validate every commit to reduce MTTR and achieve full coverage.

🔍 Experts at Infosecurity Europe warned that healthcare organizations must adopt AI-powered security to detect and contain threats faster. Legacy devices, hyper-connectivity and alert fatigue are creating a high-risk environment where ransomware and other attacks can endanger patient safety. Speakers urged proactive measures including full device visibility, clinical-risk-based prioritization, AI-driven signal correlation and segmentation to reduce exposure.

🔎 The Microsoft AI Red Team released a v2.0 update to the Taxonomy of Failure Modes in Agentic AI Systems, grounded in twelve months of red team engagements and operational data. The revision adds seven new failure mode categories—such as agentic supply chain compromise, goal hijacking, and visual attacks against computer-use agents—expands mitigations, and emphasizes supply chain, zero‑trust, and session hardening.

🔎 OpenAI proposed mandatory federal evaluations for the most capable AI models before public release while arguing regulators should not have authority to approve or block deployments. The company urged pre-release assessments by the Center for AI Standards and Innovation (CAISI) alongside audits, transparency reports, incident reporting, and whistleblower protections. OpenAI framed this approach as a middle ground that enhances government visibility and preserves developer responsibility for release decisions.

🔒 Microsoft’s Detection and Response Team (DART) warns that AI adoption has introduced new attack surfaces, with threat actors weaponizing AI tools in social engineering and supply chains. A highlighted campaign, ‘JustAskJacky’, disguised a malicious AI assistant that installed a Java backdoor and persistence tasks. Experts urge organisations to assess nonstandard AI apps, enforce security reviews, and make AI risk a board-level priority.

🔍 At Infosecurity Europe 2026, Bugcrowd unveiled ExploitBench, a graded benchmark assessing AI models' ability to chain vulnerability discovery into staged exploits against a vulnerable V8 build. Anthropic’s Claude Mythos outperformed OpenAI’s GPT‑5.5 in head‑to‑head runs, achieving higher average scores and more top‑tier exploits, often with occasional human nudges. The report highlights rising offensive potential of frontier LLMs and urges defenders to adopt automated remediation and prioritization.

🔐 At Infosecurity Europe, Forescout's Rik Ferguson urged organisations to accelerate transition plans to post-quantum cryptography (PQC), warning that only 8% of SSH servers support PQC. He cited NSA warnings about harvest-now-decrypt-later (HNDL) attacks and evidence from surveillance programs indicating encrypted data is already being hoovered for future decryption. Ferguson recommended urgent inventorying of encrypted assets, embedding quantum readiness into procurement, and building crypto-agility such as adopting TLS 1.3.

🛠️ The Amazon Bedrock console has been redesigned to match real-world model development workflows: experiment, iterate, and scale. The refreshed UI centers on the bedrock-mantle endpoint and is compatible with the OpenAI Responses API, OpenAI Chat Completions API, and the Anthropic Messages API. Users can browse and compare models, create projects to run evaluations, and get project-aware code snippets prefilled with model ID, region, endpoint URL, and API key references. The new experience is available in all Regions where the bedrock-mantle endpoint is offered.

🛡️ Security experts at Infosecurity Europe warned that expanding access to frontier AI tools for vulnerability discovery — notably Anthropic’s Project Glasswing and OpenAI’s reported GPT-5.5 Cyber pilot — heralds a structural shift in cybersecurity. Speakers advised organisations to harden controls, run incident response exercises, and accelerate adoption to avoid falling behind attackers. The panel stressed that AI augments, not replaces, human expertise; combined use improves validation and remediation of AI-discovered issues.

🚀 This tutorial shows how to build an AI agent with Google’s Agent Development Kit (ADK), containerize it, and deploy it to GKE Autopilot using Vertex AI (Gemini) as the model backend. It walks through local testing, creating a multi-stage Docker image, pushing to Artifact Registry, and configuring a Kubernetes Deployment and Service. The guide emphasizes secure authentication with Workload Identity and exposes the agent via the Kubernetes Gateway API with a Google-managed TLS certificate.

🧭 Amazon SageMaker AI introduces multi-turn reinforcement learning (RL), a serverless model customization method for fine-tuning models on multi-step, agentic tasks. The feature trains models against users' agent environments, rewarding entire decision sequences to improve task accuracy of smaller, cost‑effective models versus larger general-purpose models. It integrates with Amazon Bedrock AgentCore Runtime and other deployment targets, and handles rollout orchestration, trajectory collection, training, and checkpoints, with MLflow tracking and evaluation metrics. Multi-turn RL runs serverlessly and is available in SageMaker Studio and the SageMaker Python SDK, supporting several foundation models in specific regions.