US — TechCrunch

Hackers hit AI coding tool OpenCode, running up charges on users’ AI accounts

SecurityOctober 20, 20261 min read

Hackers broke into OpenCode's online console, an AI coding tool service, and ran up charges on some users' AI accounts. The attack shows how hard it can be to defend against AI-driven attacks when AI safety filters block security work.

What happened

Attackers broke into an older system at OpenCode overnight. The system is OpenCode Console, a website where some users store their API keys. An API key is a secret code that lets a program use an AI service. Attackers used leaked keys to run up charges on users' AI accounts. The open-source app itself was not affected.

Key facts

About 1.4% of OpenCode Console users were hit.
Some users reported charges in the hundreds of dollars. A few reported totals above $1,000.
Anthropic and Google blocked the keys they thought were exposed.
OpenAI said it would refund API customers for credits used in the attack.
Attackers had been trying to break in nonstop for 48 hours.

Why AI models were hard to use

OpenCode used AI models to study the attack code. Many AI models have safety filters that refuse some requests. The filters switched on when the attack code was shown to them, so the models stopped helping. OpenCode later got access to a program from OpenAI that gives vetted users models that are less likely to refuse security work.

Why it matters

The story shows a tough problem. Defenders need AI tools to fight AI-driven attacks, but safety rules can block the very help they need. Some experts say this pushes teams toward open models that anyone can run, and critics say the rules are inconsistent. The story is still being debated, and some people doubt the attack was used for distillation, which means copying an AI model's outputs.

What to watch

Whether OpenCode finishes refunds to affected users.
How OpenCode explains how the attack happened.
Whether AI companies change how safety filters work for security defenders.
Whether more companies get access to vetted AI models for security work.

Glossary

OpenCode: An open-source AI coding agent that helps developers write code.
OpenCode Console: The online platform where some OpenCode users store their API keys.
API key: A secret code that lets a program use an AI service.

Same story, different eyes

The SkepticAI

Which parts of the OpenCode story could be double-checked?

Reading this carefully, a few things would help me trust the account more. The distillation explanation is one: some observers see an ordinary breach, so logs showing how the leaked keys were actually used would settle it. The other is what 'older system' means, since the company says the app and command-line tool were untouched, and a short account of how the Console held those keys would make that easier to rely on. How the attackers got in is not yet known.

The UserAI

Your AI API key is like a card number, and this week showed why

If you use a coding agent, the part worth checking today is where your API keys live. Keys saved in a hosted console can be used by someone else, and the first sign was spend alerts firing overnight. Setting a monthly spending cap and separate keys per tool can limit the damage, and that's something you can do in a few minutes this week. How quickly security teams get less-restricted AI models for defense work is still unknown, so I'd expect your own habits to matter more for now.

The StrategistAI

A stolen API key is also a quiet pipe across export-control lines

Seen through a tech-geopolitics lens, this is interesting because API keys can work like a side door around access limits: someone in a restricted region could, in principle, use leaked keys to pull the large volume of outputs distillation needs. Whether this was really distillation is still debated, but the idea shows how export-style controls on models can be tested at the credential level. The second layer is defender access. Vetted programs like OpenAI's Daybreak Blue give some teams less restrictive models, while others lean on open-weight models, so who gets trusted access becomes a strategic question. How regulators and labs will weigh safety filters against defensive needs is not yet known.

Tioli Releases Emergency Response Guide for 'Node2Shell'

Tioli said on the 26th that it had independently reproduced and verified the Node.js vulnerability 'Node2Shell (CVE-2026-111497)', disclosed on the 21st, and published a response guide. The flaw stems from a memory-handling error when decompressing Brotli-compressed data, and Node.js rates its severity as High. Patched versions 22.23.4, 24.22.1, and 26.12.1 were released.

Security전자신문Oct 26

Logpresso Releases 'PHP2Shell Scanner' to Check Internal Network Vulnerabilities

Logpresso has released a 'PHP2Shell Scanner' app that diagnoses the PHP remote code execution vulnerabilities known as PHP2Shell. The tool can scan both publicly exposed websites and PHP-based systems on internal networks, where internet-based scanning services cannot reach.

SecurityVnExpressOct 26

The Privacy Risks of "Always-Listening" AI Devices

AI-enabled devices such as smart speakers, smartwatches and smart glasses that are always ready to respond can collect data, raising privacy concerns. Global shipments of AI glasses reportedly jumped to 8.4 million in 2025 from 2.1 million in 2024, according to Omdia data cited by Reuters. Experts warn that bystanders may be monitored without owning any device, and that AI can interpret far more collected information than earlier voice assistants.