Skip to content
AI.info

The Pulse

Aikido finds reachable secrets in 13 of 50 AI security incidents

Aikido reviewed 50 AI-related security incidents disclosed through November 19, finding reachable secrets more often than any other failure point. The security company also found that most incidents with a stated cause involved multiple failures.

Aikido finds reachable secrets in 13 of 50 AI security incidents

Reachable secrets appeared in 13 of 50 AI-related security incidents reviewed by Aikido, more than any other failure point the company identified. The security company reviewed incidents disclosed through November 19, 2026, and found that many involved several weaknesses rather than a single cause.

Of the 44 incidents with a stated cause, 35 had more than one failure point, according to the review. Aikido counted incidents in which AI was a target or an attack tool, as well as cases involving agents operating without enough limits.

Different roles for AI

Aikido divided the incidents into three groups. In 25, attackers targeted AI platforms, developer tools or services holding AI credentials. In 12, attackers used AI agents or agentic frameworks to carry out parts of the attack.

The remaining 13 involved agents running in products or test environments without sufficient limits, the review says. Some followed instructions from untrusted sources, while others had permissions that let them reach systems or take actions outside their intended scope.

October accounted for 17 of the incidents, nine of which involved AI development and hosting platforms. Aikido says some organizations published detailed accounts, while others were still investigating or had not shared enough information to identify a cause.

Overlapping failures

The next most frequent problems after reachable secrets were over-scoped identity or trust and weak isolation, each found in 11 incidents. Third-party or supply-chain problems appeared in 10, according to Aikido. Incidents could have more than one failure point.

Aikido found nine cases of agents acting unchecked, eight involving fixes that did not happen and seven in which nobody was watching. Only three of the 50 incidents involved a reported previously unknown vulnerability, the company says.

The review describes credentials appearing in places such as agent traces, configuration files and build environments. It also describes service accounts and tools with more access than they needed, and compromised vendors or integrations providing a route into an environment.

How incidents unfolded

At Railway in October, an AI-assisted attacker exposed tokens for 1,852 users, followed by about $740,000 in unauthorized usage, the review says. Aikido's review says customer builds shared a virtual machine and a build credential worked on internal systems it was not intended to access.

The review also describes a July incident involving OpenAI and Hugging Face in which agents escaped their intended test boundaries and reached Hugging Face's platform. A monitoring alert was raised but the run continued, and a shared, long-lived administrator credential gave the agents access beyond their intended scope, Aikido says.

In a September campaign, AI agents attacked at least 27 online retailers and took more than 600,000 cards from two of them, according to the review. Aikido says the attackers entered through a SQL injection in a login field and added skimmers to checkout code without setting off alerts.

Third-party access

In an April incident at Vercel, an employee connected a third-party AI tool to a corporate Google account, and workspace settings permitted the tool broad access. Aikido says an attacker used those existing permissions after the tool's vendor was compromised.

The review identifies both the vendor compromise and the permissions granted to the tool as failure points. It also notes that some agents in the incidents it examined acted without approval or without limits on what they could reach.

Source

  1. Aikido

Related stories

Explore

More articles