Autonomous AI Agent Attack Campaigns and Major Security Vulnerabilities Disclosed
Examining recent AI security trends, including an Amazon Threat Intelligence report on threat actors exploiting an open-source AI pentesting framework, AWS's ne
Amazon Threat Intelligence has reported a campaign in which a Chinese-nexus threat actor leveraged Cairn, an open-source autonomous artificial intelligence penetration-testing framework, to target East Asian university IP addresses over a 48-hour window, successfully exploiting a zero-day vulnerability with minimal human intervention.

Image source: https://aisecwatch.com/archive
This incident underscores the rising reality of autonomous AI agents being weaponized for sophisticated cyberattacks, arriving alongside a wave of newly disclosed high-severity security flaws and defensive patches across cloud infrastructure, open-source chat interfaces, and developer tooling.
Autonomous AI Agent Attack via Open-Source Pentesting Frameworks
According to the Amazon Threat Intelligence report, the threat actor initialized an autonomous agent given the explicit goal to "obtain shell access." Over approximately 48 hours, the agent autonomously scanned targets, identified a previously unknown and unpatched flaw—a zero-day vulnerability—in an open-source application, and took full control of the compromised server.
Operating with minimal human oversight, the agent demonstrated that autonomous AI tooling can chain reconnaissance and exploitation steps effectively, turning general-purpose testing frameworks into active cyber weapons.
AWS Permission-Narrowing API and LibreChat Patch Rollouts
In response to expanding agent-driven vectors, cloud providers and maintainers are rolling out targeted defensive measures.
Amazon Web Services (AWS) introduced a new Security Token Service (STS) API named GetScopedCredentials. This API enables agents and sub-agents to derive temporary credentials limited strictly to the intersection of the session's permissions and a supplied scope-down policy. Because permissions can only shrink and the session duration cannot exceed the parent session, delegating tasks to sub-agents no longer requires handing over full administrative privileges.
Concurrently, critical vulnerabilities affecting the LibreChat platform have been publicly disclosed and patched:
- CVE-2026-54023 (High, fixed in 0.8.6): A flaw where a GET API request returns OAuth access tokens, ID tokens, and refresh tokens directly to browser-side scripts.
- CVE-2026-54044 (High, fixed in 0.8.7): Allows a user with edit permissions on an agent to redirect an action to an attacker-controlled host, exposing stored API keys.
- CVE-2026-54039 (High, versions 0.7.0 to 0.8.6): Allows an unauthenticated attacker to forge a social-login provider value for a known victim email, gaining persistent account access and potentially administrator privileges on fresh deployments.
Memory Corruption and Command Injection Risks in MCP Tooling
The Model Context Protocol (MCP) ecosystem—connecting AI models with local environments and external tools—has also faced scrutiny following multiple vulnerability disclosures.
In codebase-memory-mcp, stack overflow risks from oversized query projections and heap corruption via crafted repository artifacts (CVE-2026-114494 and CVE-2026-114495) were patched in version 0.9.0. Indexing a malicious repository could crash the tool or potentially lead to remote code execution.
Additionally, mcp-ssh-manager (prior to version 3.8.5) was found vulnerable to shell command injection (CVE-2026-129995), permitting operating system command smuggling through unsanitized arguments. This flaw is reachable via indirect prompt injection, where malicious instructions are embedded within content read by an AI agent, and was compounded by separate tooling failing to respect read-only policy modes.
Sources
- AISecWatch Archive: Autonomous AI Agents Find and Exploit a Zero-Day