All news in category “AI and Security Pulse”

🔒 Enterprise security must stop reflexively blocking tools and start governing browser sessions. The article argues that legacy controls—kernel‑hooking agents, SSL inspection and URL blocks—create a "Workaround Economy" where users move sensitive data to unmanaged apps and extensions. It urges a shift to session‑level governance with prompt‑level DLP, extension risk scoring, and agentless clipboard/upload controls to enable productivity while reducing blind spots.

🔍 CISOs should treat AI outputs as drafts, keep humans in the loop for high‑stakes decisions, and demand traceability from vendors before accepting compliance or control assessments. The story cites practitioners who stress-test models for consistency, measure hallucination and drift rates over time, and validate AI findings against scanners and penetration testing. It warns against automated regulatory mapping without technical verification and emphasizes audit trails, human signoff, and vendor proof as essential controls.

🚨 Anthropic says it accidentally published the closed-source Claude Code source when an NPM release (v2.1.88) included a 60MB cli.js.map file that embedded original sources. The reconstructed tree contains roughly 1,900 files and 500,000 lines of code, and the leak has spread across GitHub and other platforms. Anthropic confirmed no customer data or credentials were exposed, called the incident a packaging error caused by human mistake, and is issuing DMCA takedowns while rolling out measures to prevent recurrence.

🛡️ Treat AI like a very new, junior employee and as software: it’s capable but not infallible, so give clear goals, explicit permissions, and limit its authority. Apply distinct identities and least-privilege controls, avoid relying on AI for deterministic access decisions, and test for indirect prompt injection (XPIA) using techniques such as Spotlighting and Prompt Shield. Design end-to-end systems that include people and processes, document safety plans and failure modes, and continuously monitor and vet models and agents for changes.

🛡️ AI agents are shifting enterprise automation from passive assistants to autonomous actors, creating new security challenges centered on access, autonomy, and identity governance. The article groups agents into three types—agentic chatbots, local agents, and production agents—and outlines how each carries distinct operational capabilities and risk profiles. For CISOs, the immediate priority is discovering and governing agent identities, limiting over-permissioned access, and aligning permissions with an agent’s intended purpose.

🔒 Unit 42 disclosed a permissions flaw in Vertex AI where the default Per-Project, Per-Product Service Agent (P4SA) can expose credentials and OAuth scopes via the metadata service. Researchers showed attackers could use those credentials to pivot into customer projects, read Google Cloud Storage buckets, and download images from restricted Artifact Registry repositories. Google updated docs and advises using BYOSA and least-privilege scopes; organizations should validate agent permissions before deployment.

🔒 The weaponization of AI is compressing the attack lifecycle and outpacing traditional defenses. Platforms like PlexTrac consolidate cloud misconfigurations, identity risks, application flaws, and pentest findings into a unified, dynamic view of exposure. Combined with Agentic AI for continuous threat assessment and automated remediation, organizations can prioritize actionable risk, orchestrate fixes, and validate controls at machine speed.

🔒 Unit 42 researchers discovered that AI agents deployed with Google Cloud’s Vertex AI ADK can inherit overly broad default permissions, enabling a deployed agent to leak service‑agent credentials and act as a “double agent.” By exploiting the Per‑Project, Per‑Product Service Agent (P4SA), the team pivoted into consumer projects and downloaded restricted Artifact Registry images from Google‑managed producer projects. Google collaborated with Unit 42, updated documentation, and recommended Bring Your Own Service Account (BYOSA) as a mitigation. Palo Alto Networks highlights protection via Prisma AIRS, Cortex Cloud Identity Security, and Cortex AI‑SPM.

🔐 This post summarizes the OWASP Top 10 for Agentic Applications (2026) and explains how Microsoft applies practical mitigations using Copilot Studio and Agent 365. It highlights that agentic systems merge application, identity, and data risk and can act autonomously across workflows, amplifying the consequences of failures. The article lists ten failure modes — including goal hijack, tool misuse, identity abuse, memory poisoning, and rogue agents — and outlines development and operational controls such as containment, scoped permissions, observability, and lifecycle governance to reduce exploitation and cascading impact.

🔐 At RSAC 2026, Google Cloud leaders outlined a three-stage AI adoption journey—automate tasks, redesign workflows, and rethink functions—and stressed the need for a bilingual workforce fluent in both domain and AI. They warned that AI expands the attack surface across models, agents, and data, urging multi-model, multicloud resilience and identity-centric defenses. Google highlighted the Secure AI Framework, partnerships to counter supply-chain threats like OpenClaw, and agentic SOC innovations, including the acquisition of Wiz and its AI-Application Protection Platform.

🔍 Gartner's new guidance outlines seven focused questions security teams should ask when evaluating AI SOC agents, urging outcome-driven assessments rather than feature demos. The research highlights the need to measure improvements in TDIR and MTTC, assess vendor viability and pricing, verify deep integrations with SIEM/EDR/SOAR/identity stacks, and confirm that agents transparently augment analyst skills rather than merely shifting workload. Prophet Security is cited as an example of a platform emphasizing explainable investigations and non-centralized integrations.

🔒 IronCurtain is an open-source prototype from researcher Niels Provos that confines AI agents inside isolated virtual machines and enforces user-defined security policies translated from plain English into formal rules. The approach separates agent actions from a user’s real accounts to limit access to sensitive data and reduce the impact of rogue behavior. While the containment model and interactive policy refinement are promising, the project is resource-intensive and unproven against prompt injection and other LLM-specific threats.

🔒 New research exposed a previously unknown vulnerability that allowed silent data leakage from ChatGPT conversations, challenging assumptions about AI assistants as secure containers. OpenAI has since fully resolved the flaw, but the incident underscores that enterprises must not assume AI vendors or platforms are secure by default. Security teams should validate vendor claims and apply controls before entrusting sensitive data.

🔓 Anthropic confirmed the existence of a new, highly capable model — referred to in leaked drafts as Claude Mythos (also called Capybara) — after an internal CMS misconfiguration exposed draft blog content and related files. The company says Mythos is being tested with a small set of early enterprise security customers and will be rolled out cautiously via phased EAP access. Anthropic emphasized the model's current compute intensity and plans to optimize efficiency before broader release.

🤖 Enterprise GRC teams often have the technical capability to deploy agentic AI but stall over a deeper concern: identity and role. Agents can replace operational tasks—evidence gathering, control testing, remediation tracking—but they still require human-defined logic for risk appetite, remediation criteria, and context. Anecdotes builds agentic GRC that automates operations while relying on practitioner judgment. The outcome is an opportunity for practitioners to reclaim time to focus on true risk management rather than program maintenance.

🔒 This post explains how financial institutions should augment traditional security frameworks with AI-specific controls when deploying agentic AI. It emphasizes two foundational capabilities—comprehensive observability of agent workflows and fine-grained tool access controls—to preserve explainability and accountability. The author presents seven design principles and actionable implementation guidance, referencing SR 11-7 and practical AWS tooling such as Amazon Bedrock AgentCore and monitoring integrations.

🔒 Shadow AI — the unapproved use of AI tools and embedded AI features — is proliferating as employees seek productivity gains and vendors quietly enable capabilities. CISOs should first assess data sensitivity, storage practices and whether corporate inputs are being used to train models. After evaluating risk, organizations must choose to block or formally integrate tools and apply mitigations such as filtering, acceptable-use policies and targeted employee education. Clear governance, cross-functional review and simple approval pathways help balance innovation with security without unduly punishing productive behavior.

📚 Students at UC Berkeley describe AI as a learning partner—using it to explain concepts, summarize papers, and debug code rather than as a shortcut to finished assignments. In mixed-methods interviews they framed AI as a "tutor" that extends office hours, supports students with learning disabilities, and scaffolds exploration while preserving ownership of learning. They also set explicit guardrails—limiting model access, alternating assisted and unassisted work, and asking for hints instead of full answers. This selective approach aligns with DORA findings that targeted AI use frees developers to focus on higher-level problem solving.

⚠️ Georgia Tech researchers warn that AI-assisted 'vibe coding' is producing measurable security flaws in real projects. The Vibe Security Radar traced at least 35 new CVEs in March 2026 and reports 74 confirmed AI-related vulnerabilities to date, while estimating the true count in open source may be five to ten times higher. The team monitors roughly 50 tools and uses metadata and AI agents to map vulnerable commits back to assistants such as Claude Code, noting some tools leave no trace.

🛡️ OpenAI has launched a new Safety Bug Bounty, hosted on Bugcrowd, to solicit researcher reports of AI abuse and safety risks across its products. Announced March 26, it complements the existing Security Bug Bounty and targets issues like agentic risks (MCP abuse, prompt injection, data exfiltration), account integrity violations, and proprietary-information exposures. OpenAI clarified scope limits, excludes low-impact jailbreaks, runs private campaigns for certain harms, and will triage submissions between safety and security programs.