LORD OF CLAUDE
AI NewsMCP ServersClaude SkillsEventsGuides
Subscribe to NewsletterAdvertise with Us

Autonomous Cairn AI Agents Uncover Zero-Day Vulnerability Unassisted

Amazon Threat Intelligence revealed that autonomous AI agents running the open-source Cairn framework discovered and exploited a zero-day vulnerability without.

By Betzalel Mishkan · Reviewed by Ziggy Evangeline
· 2 min read
f𝕏in✉
Illustration for the article: Autonomous Cairn AI Agents Uncover Zero-Day Vulnerability Unassisted

On September 10, 2026, Amazon Threat Intelligence discovered a publicly exposed server hosting Cairn, an open-source autonomous AI penetration‑testing framework, operating without authentication (AWS Security Blog). The server showed operational histories, prompts written in Simplified Chinese, and script execution timestamps set to China Standard Time. Analysis linked the setup to a Chinese‑nexus threat actor targeting IP addresses of East Asian universities outside mainland China during a 48‑hour campaign that began on September 8, 2026.

The operator fed the system targeted IPs and plain‑language instructions to obtain shell access. The AI agents then planned, executed, and adapted their strategy without manual steps. In one attack against a university’s unauthenticated Redis server, the agents sustained a 16‑hour assault comprising 148 sequential steps, loading unauthorized code via Redis replication, escaping the scripting sandbox, corrupting memory, and attempting Log4Shell injections.

When primary vectors failed, the agents exploited a server‑side request forgery flaw in a co‑located Qdrant instance to map internal Docker network topology. They followed up with credential‑spray attacks against PostgreSQL, SSH, and other internal services.

During the campaign, the agents uncovered a previously unknown zero‑day path traversal flaw in MONAI Label, an open‑source framework tracked via GitHub issue 1915 (GitHub). The issue lay in the PUT /datastore/image and PUT /datastore/label endpoints, where input parameters were used directly to build file paths without validation. By supplying strings such as ../../../tmp/x to LocalDatastore.add_image(), an unauthenticated remote attacker could write arbitrary files to any directory accessible by the MONAI Label process. The flaw has since been patched by enforcing path validation within the designated datastore directory.

Beyond state‑sponsored activity, autonomous agent frameworks are increasingly appearing in financial campaigns (Gambit). Gambit Security reported that a threat actor has used three open‑source AI harnesses since July 2026 to target hundreds of online retailers at a marginal cost of tens of dollars per target. Between September 10 and September 15, that campaign launched 105 projects, compromising at least 27 environments and exfiltrating over 600,000 credit‑card records.

Security research posted by a16z shows the operational window for vulnerability management is shrinking dramatically (a16z on X). a16z notes roughly 87 percent of exploited software bugs are now targeted on or before the day of public disclosure, up from 23 percent in 2020.

The autonomous discovery of zero‑day vulnerabilities by AI agents marks a operational shift in offensive cyber capabilities. By pairing goal‑oriented agents with continuous execution frameworks like Cairn, threat actors can bypass traditional operational bottlenecks. Defenders must prepare for automated, high‑velocity campaigns that systematically identify unpatched logic flaws and zero‑days faster than traditional manual security teams can issue patches.

Spotted an error? See our corrections policy or email editorial@lordofclaude.com.

ABOUT THE AUTHOR
Betzalel Mishkan, Managing Editor

Managing Editor at Lord of Claude. He oversees editorial operations and fact-checking standards, with expertise in AI security and governance, editorial leadership, technology journalism and enterprise AI.

0 Claps
Tap to clap

Discussion & Comments (0)

Join the conversation
🔒 Submitting this form adds your email to our subscriber list.

Related reading

Newsletter

New MCP servers, weekly.