All news in category “AI and Security Pulse”

🔍 Anthropic has launched Project Glasswing, an initiative that uses Claude Mythos Preview to autonomously locate and remediate undiscovered cybersecurity vulnerabilities in critical software. The private model — described by Anthropic as highly capable for coding and agentic tasks — was tested with launch partners including AWS, Google and Microsoft and reportedly found thousands of previously unidentified zero-day flaws. Anthropic committed up to $100m in usage credits and $4m in donations to support open-source security while keeping Mythos Preview restricted to defenders with guardrails.

🔐 Two independent research efforts from Irregular and Kaspersky demonstrate that modern LLMs produce passwords that are structurally predictable and far lower in effective entropy than they appear. Models often repeat the same strings across sessions and conform to human-like patterns that fool standard strength meters. Autonomous coding agents are embedding these credentials into configuration files and repositories, and conventional secret scanners lack the means to detect them. Organizations should audit codebases, rotate suspect credentials, and require explicit use of cryptographically secure RNGs for all generated secrets.

🔒 Agentic AI is accelerating vulnerability discovery and shrinking the window between unknown flaws and active exploitation. A zero-day is dangerous because it exists in a defensive vacuum with no vendor patch and no established playbook, forcing emergency responses. Automated agents can probe, adapt and iterate continuously, so periodic assurance measures like quarterly scans and annual penetration tests are no longer sufficient as the primary resilience strategy. Organizations should emphasize data minimization, strict API discipline, least-privilege controls and micro-segmentation while embedding security into day-to-day IT operations and aligning CIO and CISO responsibilities.

🛡️ Microsoft has released the open-source Agent Governance Toolkit to monitor and control AI agents during runtime as organizations move them into production. The toolkit enforces policies aligned with OWASP top risks for agentic systems, such as prompt injection, identity abuse, and tool misuse, while improving visibility across multi-step workflows. It ships as multi-language components and integrates with existing frameworks like LangChain without requiring agent rewrites. The project is in public preview under an MIT license.

🔐 Anthropic launched Project Glasswing to apply a preview of its frontier model, Claude Mythos, to find and help remediate security vulnerabilities in critical software. The company says Mythos Preview has already identified thousands of high‑severity zero‑day flaws and autonomously developed complex exploits in testing. Access is restricted to a small set of vendors and foundations due to abuse risks. Anthropic committed significant usage credits and donations to support coordinated defensive patching while acknowledging prior operational leaks and the risk that the same capabilities could be misused.

🤖 Traditional tabletop exercises build shared understanding, clarify escalation paths and satisfy compliance, but they often test knowledge of a plan rather than the ability to execute it. The authors—experienced facilitators—note scripted injects and calls to “suspend disbelief” reveal a gap between documentation and operational reality. AI agentic capabilities can simulate adaptive adversaries and reactive stakeholders, turning static scenarios into dynamic, consequence-driven drills. Even so, skilled facilitators and a judgment-focused post-mortem remain essential.

🔎 Anthropic's Project Glasswing uses Claude Mythos Preview to autonomously hunt software vulnerabilities and is being offered to a closed consortium of more than 40 organizations, including Amazon, Microsoft, Apple, Google and the Linux Foundation. Anthropic says early tests found thousands of high-severity flaws across operating systems, browsers, and other widely used software, including an allegedly 27-year-old OpenBSD bug. Security leaders warn the development could upend bug-bounty economics, push security upstream, shorten exposure windows, and raise dual-use control questions.

🔐 AI is rapidly changing how software is produced, introducing a new class of instant software that is written, deployed, and discarded on demand. This shift alters vulnerability dynamics because AIs can both discover and craft exploits as well as generate patches, empowering attackers and defenders simultaneously. The balance of power will hinge on how quickly AIs learn to write secure code, reliably produce updates, and coordinate defensive sharing.

🔒 Google explains how it designed Gmail to protect user data as Gemini-powered features roll out. The company says Gemini is not trained on personal email content and only accesses messages for specific, isolated tasks like summarization. According to Gmail’s VP of product, Blake Barnes, the feature processes requests inside the inbox and does not retain the processed data.

🔒 Attackers are compressing the time from initial access to lateral movement by using AI, automation and refined TTPs, forcing defenders to adopt prevention-first strategies. The article highlights that average breakout time is about 30 minutes and that exfiltration can sometimes occur in minutes, with extreme cases measured in under ten minutes. It recommends AI-powered XDR/MDR, unified visibility across endpoint, network and cloud, and stronger identity-centric controls to speed detection and response. Automated containment—session termination, host isolation and password reset—should be orchestrated with SIEM and SOAR to reduce dwell time.

🔒 CrowdStrike announced it is a founding member of Project Glasswing, partnering with Anthropic to secure execution of frontier models like Mythos Preview where they run inside enterprises. CrowdStrike emphasizes its sensor-level visibility across endpoints, real-time AI Detection and Response, and Falcon Data Security to govern data and agent behavior at runtime. The company frames deployment governance as distinct from model safety and highlights regulatory and operational requirements for enterprise adoption.

⚠️ Attackers are increasingly abusing enterprise AI services—poisoning connectors, impersonating Model Context Protocol (MCP) servers, and using platforms as covert C2 channels—to exfiltrate sensitive data and hide malicious traffic. Notable incidents include a counterfeit MCP package siphoning transactional emails, the SesameOp backdoor tunneling commands through the OpenAI Assistants API, and command-injection flaws in Microsoft Copilot and OpenClaw that enabled agent hijacking. Threat actors also automate espionage with Claude Code and assemble modular black‑hat stacks like Xanthorox and Hexstrike. Security teams should treat AI assistants like privileged users, enforce governance, and harden supply-chain and connector integrity.

🔒 This Unit 42 analysis evaluates Amazon Bedrock Agents' multi-agent collaboration from a red-team perspective. The researchers demonstrate a chain of reconnaissance and exploitation—detecting operating mode, enumerating collaborator agents, delivering attacker-controlled payloads, and triggering tool actions—when Bedrock Guardrails and pre-processing are disabled. The report confirms no vulnerabilities in Bedrock itself and emphasizes mitigations such as Bedrock Guardrails, input validation, scoped agent capabilities, and the principle of least privilege.

🛑 The Internet Bug Bounty program, administered by HackerOne and backed by multiple major software companies, has paused submissions and payouts while it reassesses how best to support open source security. HackerOne said the rise of AI-assisted vulnerability discovery has increased both coverage and speed, shifting the balance between new findings and remediation capacity. Projects such as Node.js will continue to accept and triage reports via HackerOne but may not issue rewards from the paused fund. Similar changes have hit other programs, including curl and recent restrictions at Google's open source rewards effort.

🔒 A leaked copy of Claude Code has revealed a documented vulnerability that can be triggered when the tool receives more than 50 subcommands. Researchers at Adversa found that subcommands beyond the 50th bypass compute-intensive security analysis and instead elicit a simple user confirmation, creating a risky blind spot. Anthropic has developed a fix — a tree-sitter parser — but it is present only in internal code and not enabled in public builds that customers use.

🔒AWS outlines four security principles for agentic AI in its NIST CAISI response, arguing existing security frameworks should be extended rather than replaced. It emphasizes secure development lifecycles for both traditional and AI components, continued use of standard controls, and deterministic, infrastructure-level enforcement outside the agent's reasoning ('security box'). AWS applies these through Amazon Bedrock AgentCore, which provides compute isolation, identity and access controls, centralized tool gateways, observability, and secure model execution.

🔐 Google outlines a continuous, layered approach to mitigating indirect prompt injection (IPI) across Workspace with Gemini, combining proactive discovery, synthetic data generation, and iterative defenses. Human and automated red-teaming, an AI Vulnerability Rewards Program, and OSINT monitoring are used to catalog and expand attack variants. Deterministic configuration controls, ML retraining, LLM prompt hardening, and model-level defenses are validated through comparative testing to reduce IPI success while preserving routine performance.

🤖 Microsoft describes a shift from AI as a tool to AI as an embedded attack surface, accelerating tempo, precision, and scale across reconnaissance, malware development, and post-compromise activity. AI-enhanced phishing campaigns now report click-through rates near 54% versus roughly 12% for traditional campaigns, a 450% increase. The blog highlights Tycoon2FA, tied to Storm-1747, as an industrialized, subscription-based phishing ecosystem that automated MFA bypass at scale. Microsoft’s Digital Crimes Unit disrupted the operation, seizing 330 domains with Europol and partners, and urges organizations to prioritize agent inventory, agentic accountability, and lifecycle-integrated intelligence and defenses.

🔐 AI is rapidly reshaping how software is written, deployed, and consumed, pointing toward a future of on-demand "instant software" that is created and discarded as needed. The essay examines how improved AI tools will change the attacker/defender dynamic by automating both vulnerability discovery and, potentially, patch creation. It highlights particularly exposed areas such as IoT and legacy industrial systems and outlines several key unknowns—AI effectiveness on closed-source code, patch reliability, update lag, coordination of defenses, and risks of poisoning or social-engineering attacks. The author sketches optimistic scenarios (self-healing networks, rapid coordinated patching) while warning that attackers will adapt by targeting unpatchable legacy code and human elements.

⚠️ Last week a leaked build of Anthropic's new model, Claude Capybara (also called Mythos), revealed substantially improved capabilities for automated vulnerability discovery, exploit development, and multi-step attack reasoning. The incident marks a turning point: frontier AI can compress attack lifecycles and enable scalable, novel exploitation techniques that were once the domain of advanced state actors. Security teams should treat this as a warning and accelerate risk assessments, patching, detection, and governance measures.