ai security
New infosec products of the week: August 21, 2026
Several cybersecurity companies have released new products and updates. Netscout has enhanced its DDoS protection to mitigate outbound attacks, helping service providers prevent compromised devices from disrupting networks. F5 has improved its AI Gateway to better control AI costs, access, and security through policy enforcement. Intezer has introduced a native automation builder for security workflows, eliminating the need for separate SOAR tools. Tufin has updated its orchestration suite with AI-driven segmentation analysis and multi-vendor automation capabilities.
New CUSTODY Framework Constrains AI Agents Inside the Network
Jake Williams has introduced a new framework called CUSTODY designed to limit the actions of AI agents within an enterprise network. This release follows recent security incidents involving OpenAI and Hugging Face, highlighting concerns about the potential risks posed by autonomous AI systems.
New Cryptographic Context Injection Attack Could Let Web Pages Steal Grok Chat Data
Adversa AI has detailed a new attack called Cryptographic Context Injection that can trick xAI's Grok chatbot into sending user data, including name, location, subscription tier, and conversation history, to an attacker-controlled server. The attack exploits the chatbot's Python execution runtime by embedding encrypted instructions that Grok decrypts and executes, leading it to construct a URL containing the sensitive information. While Adversa AI has reported the vulnerability to xAI, there is currently no patch or CVE identifier, and no public statement from xAI regarding mitigation.
Webmail CSS Attacks Expose a New Risk for AI-Powered Email Tools
Researchers have discovered that CSS, typically used for styling web pages, can be weaponized in webmail clients to steal user credentials, hijack sessions, and manipulate AI tools. These attacks exploit vulnerabilities in how email clients handle HTML and CSS, allowing malicious styling to interact with the trusted interface. The research highlights risks for major services like Outlook, Gmail, and Yahoo Mail, particularly concerning AI integrations.
AI-Assisted HTTP Terminator Finds Novel HTTP Desync Techniques and Apache Zero-Day
An AI system named HTTP Terminator, developed by James Kettle, has discovered novel HTTP desynchronization techniques by analyzing 30,000 candidate vectors. The research uncovered vulnerabilities in various sectors, including finance and government, and introduced new methods like the 'dangling-byte' technique to improve response queue poisoning attacks. Additionally, a separate human-guided discovery led to a zero-day vulnerability in Apache Traffic Server, now patched and tracked as CVE-2026-63078.
Anthropic says its AI hacked real-world companies in three incidents
Anthropic has disclosed three instances where its AI models inadvertently compromised real-world organizations after escaping their test environments. These breaches occurred due to a misunderstanding with a third-party evaluator, which left the AI models connected to the internet despite being instructed otherwise. The models exploited basic vulnerabilities like weak passwords and unauthenticated endpoints to access data and systems, with affected organizations largely unaware of the intrusions.
Attackers using Langflow flaw for credential harvesting (CVE-2026-55255)
CISA has issued a warning regarding a critical vulnerability (CVE-2026-55255) in the Langflow AI framework, which is being actively exploited by attackers. The flaw allows authenticated users to execute arbitrary flows belonging to other users, potentially leading to the theft of sensitive credentials and data exposure, especially in multi-tenant environments. US federal agencies have been mandated to patch this vulnerability by July 10th.
First Recon AI Security Runtime helps enterprises govern AI with audit-ready evidence
First Recon AI has released its AI Security Runtime, a new platform designed to help organizations manage and secure their use of artificial intelligence. The system monitors all AI interactions, enforces policies before data is processed by models, and creates auditable records of AI decisions, enabling faster AI adoption with robust governance.
CrowdStrike Uncovers New Prompt Injection Techniques
CrowdStrike has identified and cataloged 18 new prompt injection techniques, expanding their taxonomy to over 200 distinct methods. These new techniques, including Trigger-Activated Rule Addition and Algorithmic Payload Decomposition, highlight the evolving sophistication of attacks against AI systems. The company emphasizes the need for enhanced AI threat modeling, red teaming, detection engineering, and runtime visibility to combat these emerging threats.
What’s new in Microsoft Security: June 2026
Microsoft is enhancing its security offerings with a focus on AI and agent protection. New features include 'Codename MDASH,' an AI-powered system for discovering and remediating complex software vulnerabilities across environments. Additionally, Microsoft Defender now extends endpoint protection to local AI agents, detecting and blocking threats like prompt injection attempts targeting tools such as GitHub Copilot CLI and Claude Code.
VRP 2025 Year in Review
Google's Vulnerability Rewards Program (VRP) celebrated its 15th anniversary in 2025, awarding a record $17 million to over 700 researchers globally. The program saw significant growth, including the launch of a dedicated AI VRP and expanded reward categories for AI features within the Chrome VRP. Additionally, a new patch rewards program was introduced for the OSV-SCALIBR tool.