AI Security Bulletin, August 2026, Published
Overview
The Security Center of the Information-technology Promotion Agency, Japan, has published the August 2026 issue of the AI Security Bulletin, which introduces AI security trends and cases based on public information released daily in Japan and abroad. Focusing on developments that AI system developers and security personnel should know and incident cases requiring attention, it is organized into four fields: ensuring safety related to AI, ensuring cybersecurity using AI, responding to cyberattacks that misuse AI, and AI safety.
Key points
- The August 2026 issue of the AI Security Bulletin is a reference document compiling AI security developments and cases published by the IPA Security Center.
- Its content is organized into four fields: ensuring safety related to AI (Security for AI), defense using AI, responding to attacks that misuse AI, and AI safety.
- The table of contents includes numerous individual items concerning vulnerabilities, AI agents, incidents, threat countermeasures, and other topics.
- The published content reflects information available at the time of each investigation, and circumstances may change or additional information may be released after publication.
Overview
This document is the August 2026 issue of the AI Security Bulletin produced by the Security Center of the Information-technology Promotion Agency, Japan (IPA). It introduces AI security trends and cases from the vast amount of AI-related public information released daily in Japan and abroad.
It focuses specifically on developments that AI system developers and security personnel should know in their daily work, as well as incident cases requiring attention. Following the executive summary, the table of contents is organized in the order of Ensuring Safety Related to AI (Security for AI), Ensuring Cybersecurity Using AI (AI for Security), Responding to Cyberattacks That Misuse AI, and AI Safety.
The executive summary raises the issues of vulnerability lifecycle N-hourization, the transformation of the entire development process into an attack surface through AI coding agents, fully autonomous threat actors, Human-in-the-Loop, applying zero trust to agentic AI, and rogue AI agents.
Impact
The bulletin is positioned as information for AI system developers and security personnel to learn about developments relevant to their daily work and incident cases requiring attention.
The published information reflects the status at the time of the investigation, and circumstances may change or additional information may be released after publication.
Details
Ensuring Safety Related to AI (Security for AI) covers ChatGPT Lockdown Mode and session management; compromise rates and identity control accompanying expanded AI use; flaws in Langflow and LangGraph; an ad-blocker extension collecting AI conversations; reverse-shell execution and private repository leaks by AI agents; approval-bypass flaws; sandbox escapes; prompt injection; and the npm worm CHAINDROP.
Ensuring Cybersecurity Using AI (AI for Security) covers Microsoft's updates; OpenAI's Daybreak extension and Patch the Planet; vulnerability coordination through GOLD EAGLE; Google's Gemini 3.5 Flash Cyber and CodeMender; Cisco's compact model; the Open Secure AI Alliance; security-specialized models; and automation of vulnerability discovery and remediation in Chrome.
Responding to Cyberattacks That Misuse AI covers the acceleration of N-day exploitation by AI, risk-based remediation deadlines, phishing networks, AI coding agents, exploitation of Langflow, JADEPUFFER, AWS compromises, autonomous attacks, and exploitation status after PoC publication. The table of contents includes an item stating that shell boundaries can be bypassed in 10 out of 11 AI coding agents and an item stating 88% within 48 hours of PoC publication.
AI Safety covers Anthropic's Fable 5 and Mythos 5; stop orders issued and lifted by the United States government; OpenAI's flagship model Sol; the Hugging Face compromise; unauthorized actions by AI agents under evaluation; cyber capability assessments of the next-generation model Astra; Kimi K3; and unauthorized intrusion into a gym reservation system.
Back issues listed are the June 2026 issue, the March 2026 issue, the December 2025 issue, the September 2025 issue, and the July 2025 issue.