Review of 50 AI Security Incidents Finds Familiar Failures Behind 2026 Cases

Summary
Aikido reviewed 50 publicly reported AI-related security incidents through November 19, 2026. Exposed secrets, excessive permissions, weak isolation and vendor compromises recurred; the sample is incomplete and some causes remain unknown.
Key points
- The review covers 50 publicly reported incidents through November 19, 2026, and is not a complete account of all AI-related incidents.
- AI was targeted in 25 incidents, used as an attack tool in 12, and involved in insufficiently constrained agent deployments in 13.
- Of 44 incidents with a stated cause, 35 had more than one failure point; only three involved a reported previously unknown vulnerability.
- Recurring issues included exposed credentials, over-scoped access, compromised vendors, weak isolation and failures to act on alerts.
- The article says agentic frameworks can automate reconnaissance and attacks, while agents can exceed intended boundaries through untrusted instructions, excessive permissions or weak isolation.
- Recommended defenses include limiting credentials and agent permissions, isolating environments, monitoring unusual activity and requiring approval for irreversible actions.
- Six incidents had no specific cause reported, and 17 reports lacked at least one detail; the authors note that their categories and counts may change as more information emerges.
Article Details
- Publisher
- Aikido
- Report Period
- 2026 through November 19
- Scope
- Review of publicly reported AI-related security incidents and their stated causes.
- Sample Size
- 50 incidents; the article notes this is not a complete picture of all AI-related incidents.
- Key Statistics
- Of 50 incidents, 25 involved attackers targeting AI platforms, developer tools, or services holding AI credentials.
- Of 50 incidents, 12 involved attackers using AI agents or agentic frameworks to carry out parts of an attack.
- Of 50 incidents, 13 involved AI agents operating without sufficient guardrails.
- Three of the 50 incidents involved a reported previously unknown vulnerability.
- Of 44 incidents with a stated cause, 35 had more than one failure point.
- Recommendations
- Redact secrets from agent logs and traces, and keep build secrets out of environments that do not need them.
- Audit connected apps, OAuth grants, and service accounts; limit access to specific resources.
- Enforce network limits for agent sandboxes and build environments rather than relying on prompts to keep agents in scope.
- Alert on unusual data volumes and API calls, including those from agents, and decide who is responsible for responding.
- Limit agent permissions to their tasks and require approval for irreversible actions.
MITRE ATT&CK
Threat Actors
Vendors
AnthropicGitHub issue titles that anyone could write, contributing to an unauthorized npm release. And in Anthropic's cybersecurity evaluations, agents reached real systems because their test environments had internetConnectWiseConnectWise (October 2026). Storm-0940 used ARTEX, an agentic offensive framework, to attempt sign-ins against more than 2,250 tenants. ARTEX automated reconnaissance, authentication attempts, and actions such asHugging FaceOpenAI and Hugging Face (July 2026). AI agents escaped their intended test boundaries and reached Hugging Face's platform. The sandbox could reach systems it shouldn't have, monitoring raised an alert but the runMetaMeta (April-May 2026). Attackers used an AI support assistant to initiate password resets for Instagram accounts they didn’t own. The recovery system failed to verify that the supplied email belonged to the account,MicrosoftWhat AI did change is the pace and the prize. In one campaign Microsoft tracked, there were multiple cases where attackers using an agent framework went from first targeting to privileged access in under three hours.OpenAIOpenAI and Hugging Face (July 2026). AI agents escaped their intended test boundaries and reached Hugging Face's platform. The sandbox could reach systems it shouldn't have, monitoring raised an alert but the runRailwayRailway (October 2026). An AI-assisted attacker exposed tokens for 1,852 users, and about $740,000 of unauthorized usage followed. The path there was customer builds sharing a VM, plus a build credential that worked onVercelThese problems often overlap. Take Vercel’s April incident. An employee connected a third-party AI tool to their corporate Google account, and the workspace settings allowed it broad permissions. When the tool’s vendor
Products
Clinetransferred funds after following instructions relayed through another AI's public reply. In February, Cline's issue-triage bot was manipulated through GitHub issue titles that anyone could write, contributing to anGitHubrelayed through another AI's public reply. In February, Cline's issue-triage bot was manipulated through GitHub issue titles that anyone could write, contributing to an unauthorized npm release. And in Anthropic'sInstagramMeta (April-May 2026). Attackers used an AI support assistant to initiate password resets for Instagram accounts they didn’t own. The recovery system failed to verify that the supplied email belonged to the account,npmbot was manipulated through GitHub issue titles that anyone could write, contributing to an unauthorized npm release. And in Anthropic's cybersecurity evaluations, agents reached real systems because their testZammadcode running on 41 production servers. At DIVD, an AI agent chained two previously unknown flaws in its Zammad helpdesk to get root access, but network segmentation kept the damage mostly to volunteer contact details.
