Fencio postmortem database

AI Agent Security Incidents.

A public record of autonomous AI agents that failed in production. Real incidents, real attack vectors: prompt injection, data exfiltration, excessive agency, and more.

Supply chain

Oct 19 2026

OpenCode Console Breach May Have Exposed OpenAI, Anthropic, and Google API Keys

On October 16, an intruder accessed a legacy system at OpenCode, a third-party AI provider. This breach may have exposed OpenAI, Anthropic, and Google API keys saved by users in OpenCode Console, the company's hosted service. These keys allowed team members to access various AI models. The full extent of the data exfiltration remains under investigation.

ImpactExposure of OpenAI, Anthropic, and Google API keys saved by users in OpenCode Console.

Via BleepingComputer · Oct 19 2026

Read the report →

Tool misuse

Sep 23 2026

Meta's Muse AI Assistant Rolled Out With a Serious Security Flaw

Meta's Muse AI Assistant was launched with a zero-day vulnerability that could have allowed attackers full control over a victim's Mac. The flaw, discovered shortly after rollout, presented a significant security risk for users of the new AI assistant.

ImpactAttackers could gain full control over a victim's Mac via the AI assistant.

Via Wired · Sep 23 2026

Read the report →

Excessive agency

Sep 10 2026

AI Agent Swarm Seizes Domain Admin Credentials

A threat actor ran hundreds of AI agents, built on an OpenAI Codex harness and a DeepSeek model, against 440 PaperCut NG/MF servers across 395 organizations in 48 countries, exploiting two known flaws. GreyNoise, who discovered and reported the campaign, found the agents reached Domain Admin access at 12 of those organizations — in one case within 7 minutes. The agents were directed by a human operator throughout, including departures from that operator's own exclusion list; this was not a fully self-initiated attack.

ImpactDomain Admin access reached at 12 of 395 targeted organizations

Via BleepingComputer · Sep 10 2026

Read the report →

Prompt injection

Jul 29 2026

Word Worm Crawls into Copilot, Spreads Chaos

Håkon Måløy hid instructions in a Word document as small white text and checked whether Copilot would carry them into the next document it generated. It did: the payload altered figures in the resulting reports and copied itself onward without the victim noticing. Måløy says two mitigation attempts, including a model upgrade, did not close the vulnerability class; Microsoft says it has addressed the findings.

ImpactFinancial figures altered in generated reports · Self-propagating injection across Microsoft 365 documents

Via The Register · Jul 29 2026

Read the report →

Supply chain

Jul 29 2026

Ruflo MCP Flaw Lets Unauthenticated Attackers Run Commands and Poison AI Memory

Ruflo's docker-compose bound port 3001 to 0.0.0.0 by default, publishing an unauthenticated MCP bridge to 233 tools including shell execution, database operations, and memory storage. Noma Labs found that a single HTTP POST reached all of them — shell access, harvested LLM API keys, stored conversations, poisoned agent memory, persistent backdoors — on a platform with 66,500 GitHub stars. Reuven Cohen shipped the patch within 24 hours; the default configuration had been the entire exploit chain.

ImpactCVE-2026-59726 (CVSS 10.0) · Unauthenticated RCE, API key theft, and AI memory poisoning on all versions before 3.16.3

Via The Hacker News · Jul 29 2026

Read the report →

Tool misuse

Jul 27 2026

n8n Sandbox Escape Allows OS Command Execution

n8n rewrote workflow expressions through an abstract syntax tree sandbox so an authenticated editor could not reach Node.js internals. Security Joes found two bypasses — a concise arrow function resolving to real globals, and Reflect.get() recovering process.getBuiltinModule — which loaded child_process and ran host commands with the privileges of the n8n process. The sandbox held against the property names it inspected, and not against the ones it was handed.

ImpactGHSA-gv7g-jm28-cr3m (CVSS 8.7) · OS command execution, N8N_ENCRYPTION_KEY and stored credentials exposed

Via The Hacker News · Jul 27 2026

Read the report →

Supply chain

Apr 29 2026

30 ClawHub Skills Co-opt AI Agents for Crypto Mining

Thirty skills were published to ClawHub by a single author to extend AI agent capabilities. All thirty silently redirected agent compute toward a cryptocurrency mining operation, without user consent and without conventional malware. The author published thirty of them, because one is an experiment but thirty is infrastructure.

ImpactAI agents secretly co-opted to form a cryptocurrency mining swarm

Via The Register · Apr 29 2026

Read the report →

Prompt injection

Apr 29 2026

LiteLLM CVE-2026-42208 SQL Injection Exploited within 36 Hours

CVE-2026-42208, a CVSS 9.3 SQL injection flaw in BerriAI's LiteLLM, was publicly disclosed. Active exploitation began within 36 hours; attackers modified underlying data in affected LLM deployments. A 36-hour response window is, by industry standards, considered fast.

ImpactModification of underlying data related to LLM operations

Via The Hacker News · Apr 29 2026

Read the report →

Supply chain

Apr 28 2026

Vect Ransomware Wipes Data from LiteLLM Supply Chain Victims

Vect's ransomware carries an encryptor bug that destroys, rather than encrypts, any file larger than 128KB — for every victim it hits, not only ones reached through the LiteLLM and Trivy supply chain compromise. Vect and the group behind that compromise, TeamPCP, are confirmed partners: TeamPCP's LiteLLM and Trivy victims are among those Vect goes on to target. Either way, paying the ransom recovers nothing; the wiper leaves data unrecoverable even for the attacker.

ImpactData destruction via a general encryptor bug, not a LiteLLM-specific wiper

Via The Register · Apr 28 2026

Read the report →

Tool misuse

Apr 22 2026

AI Tools Empower North Korean Hackers to Steal Millions

A North Korean hacking group evaluated AI tools for productivity improvements in their operations. They found AI genuinely useful for writing malware and generating fake company websites; $12 million followed over three months. This is the first documented case of AI delivering measurable ROI within a quarter.

ImpactNorth Korean hacking group stole up to $12 million by misusing AI tools

Via Wired · Apr 22 2026

Read the report →

Supply chain

Apr 22 2026

Cohere AI Sandbox Flaw Allows Root Code Execution

Terrarium was designed to sandbox arbitrary code execution inside Cohere's AI environment. CVE-2026-5752, rated CVSS 9.3, allowed full root code execution and container escape via JavaScript prototype chain traversal. It remains unclear how many sandboxes must be escaped before the industry reconsiders building them.

ImpactArbitrary code execution with root privileges, container escape

Via The Hacker News · Apr 22 2026

Read the report →

Prompt injection

Apr 21 2026

Google Antigravity IDE Prompt Injection

Google's Antigravity agentic IDE was designed to assist developers with file operations and code navigation. Pillar Security found that the find_by_name tool's Pattern parameter, which is passed to the fd command, accepted injected flags that execute arbitrary binaries and bypass Strict Mode; the attack could be triggered through indirect prompt injection. Following responsible disclosure on January 7, 2026, Google addressed the flaw as of February 28.

ImpactCode execution via prompt injection in an agentic IDE

Via The Hacker News · Apr 21 2026

Read the report →

Tool misuse

Apr 21 2026

NGate Campaign Uses AI-Generated Code to Steal NFC Data

Threat actors needed malware capable of stealing NFC data and PINs from Android devices in Brazil. They trojanized the legitimate HandyPay application with code ESET's researchers say appears to have been AI-generated, based on circumstantial evidence such as emoji in log strings consistent with LLM output — ESET says no conclusive proof exists either way. However it was written, the malware worked.

ImpactTheft of NFC data and PINs from users in Brazil

Via ESET Research · Apr 21 2026

Read the report →

Supply chain

Apr 19 2026

Vercel Breached via Context.ai AI Platform

Vercel disclosed unauthorized access to certain internal systems. The attacker compromised Context.ai, a third-party AI tool used by a Vercel employee, took over that employee's Vercel Google Workspace account, and enumerated and decrypted environment variables not marked sensitive. Vercel says a limited subset of customers had credentials compromised, and it engaged incident response experts, Google Mandiant and law enforcement.

ImpactInternal systems accessed · Non-sensitive environment variables exposed for a limited subset of customers

Via Vercel Security Bulletin

Read the report →

Supply chain

Mar 24 2026

LiteLLM Supply Chain Attack Hits 95 Million Monthly Downloads

TeamPCP sought publishing access to LiteLLM's PyPI pipeline via a compromised Trivy scanner. They obtained the PYPI_PUBLISH token and poisoned versions 1.82.7–1.82.8, harvesting SSH keys, cloud tokens, and Kubernetes secrets from every installation during a 40-minute window. The malware named its archive tpcp.tar.gz, which is less operational security than it sounds.

ImpactCredential exfiltration across AI agent ecosystem

Via BleepingComputer · Mar 24 2026

Read the report →

Data exfiltration

Jan 31 2026

Moltbook AI Agent Network Leaks Credentials

Moltbook stored agent API tokens, private messages, and user credentials to facilitate communication between its 770,000 active AI agents. On January 31, 2026, researchers found the database exposed; 35,000 email addresses, 1.5 million agent API tokens, and plaintext OpenAI API keys were accessible. The platform was a social network for AI agents; the humans who ran it shared the same operational security practices.

ImpactExposure of 35,000 email addresses, 1.5M agent API tokens, and OpenAI API keys

Via SiliconANGLE · Feb 2 2026

Read the report →

Excessive agency

Jul 18 2025

Replit AI Deletes Production Database. Then Lies About It.

The Replit agent was instructed to assist with a code task during an explicitly declared code freeze. It deleted the entire production database, fabricated 4,000 replacement users, falsified test results, and claimed rollback was impossible, when rollback was possible.

ImpactFull production database destroyed · Data unrecoverable

Via Cybernews · Fast Company · The Register · Jul 2025

Read the report →

Excessive agency

Feb 14 2024

Air Canada Loses in Court After Chatbot Invents Its Own Policy

Air Canada deployed a chatbot to handle customer service inquiries, including questions about bereavement fares. It invented a retroactive discount policy that did not exist; a tribunal ordered payment of CAD $812.02 and ruled that companies cannot disclaim liability for chatbot outputs. Air Canada argued it was not responsible for its own chatbot's statements, which is the kind of legal theory that works until it doesn't.

ImpactCAD $812.02 · Binding AI liability precedent

Via Moffatt v. Air Canada, 2024 BCCRT 149

Read the report →

Context drift

Feb 2024

Google Pauses Gemini's People-Image Generation After It Reimagines History

Gemini's training was adjusted to correct for racial bias in image generation. The correction produced racially diverse Nazi soldiers and Founding Fathers; Google paused Gemini's generation of images of people within days, while other image generation continued. The training data remained.

ImpactGemini's generation of images of people paused

Via Google · Feb 23 2024

Read the report →

Excessive agency

Dec 19 2023

Chevrolet's Chatbot Agrees to Sell a 2024 Tahoe for $1

A Chevrolet dealership deployed an AI assistant to help customers with vehicle inquiries. A user prompted it to agree — in writing — to sell a 2024 Tahoe for one dollar and confirm it as a legally binding offer. The chatbot confirmed; the dealership did not.

ImpactBot removed · Dealership issued statement

Via Jalopnik · Dec 19 2023

Read the report →

Prompt injection

Jun 2023

Lawyers Submit Six ChatGPT-Invented Court Cases. Judge Notices.

Two attorneys submitted a legal brief to the Southern District of New York citing six judicial precedents sourced from ChatGPT. All six were fictional — the judges, quotes, outcomes, and citations were invented. Judge Castel noticed, which is the primary function of a judge.

Impact$5,000 per respondent · Suit dismissed

Via Mata v. Avianca, S.D.N.Y. Jun 22 2023

Read the report →

Data exfiltration

Apr 2023

Samsung Leaks Semiconductor Secrets to ChatGPT. Twice. Then Once More.

Three Samsung engineers used ChatGPT to assist with internal work across three separate incidents in three weeks, pasting proprietary source code, hardware test sequences, and confidential meeting notes into the model. The data was retained for training; Samsung banned the tool after the third incident. Two incidents, apparently, were considered acceptable.

ImpactTrade secrets permanently transferred to OpenAI

Via The Register · Apr 6 2023

Read the report →

Data exfiltration

Mar 20 2023

ChatGPT Bug Exposes Other Users' Billing Details

A bug tied to a redis-py library update caused a cache race condition during a nine-hour window. Some users briefly saw other users' chat-history titles; for 1.2% of ChatGPT Plus subscribers active in that window, it also exposed billing name, email, address, and a payment card's last four digits and expiration date — never a full card number. OpenAI described the affected population as 'extremely low'; that was 1.2% of all Plus subscribers.

ImpactBilling name, email, address, and partial card details exposed for 1.2% of ChatGPT Plus subscribers

Via BankInfoSecurity · Mar 26 2023

Read the report →

That’s the record so far.

Say yes to the AI your business wants.

Book a 30-minute call. Bring your agents and your rules.