OpenCode said on 17 October that attackers had broken into a system behind OpenCode Console, the company’s optional hosted gateway for AI models. Its post described the incident as “a large scale distillation attack” and said it affected 1.4% of OpenCode Console users, who were notified by email. The affected group is narrower than that number suggests: OpenCode said it was “a subset of users who had external providers linked”, meaning users who had stored third-party credentials in the Console.
Console is a different product from the OpenCode app and CLI. It is a hosted layer where a team connects provider accounts it already pays for, such as Anthropic, OpenAI or AWS Bedrock, and shares those models across a workspace. Co-founder Dax Raad said the app and CLI were not affected, because “those keys never come to us”.
The keys in Console do come to OpenCode. The Console provider documentation is explicit: “Console stores the credential and serves each provider through its gateway, so members use the models without ever seeing the provider key.” Workspaces hold members, credits, budgets, providers, policies and service accounts.
Raad said attackers hammered the company’s systems for about 48 hours, as TechCrunch reported on 19 October. Developers posted spending alerts across OpenAI, Anthropic and Google accounts, and TechCrunch reported charges in the hundreds of dollars, with some above $1,000.
The providers responded differently. OpenAI said it restored access and refunded credits drawn in the attack, in a post from Tibo Sottiaux, an OpenAI employee. Anthropic revoked Claude API keys believed affected, according to Lydia Hallie of the Claude Code team. Google blocked exposed Gemini API keys and is working with affected customers, according to Evan Otero, a Google engineer posting from a personal account. Among the providers, only OpenAI has said anything about refunds. If you held a key in Console during the incident, revoking it and replacing it is the action that does not depend on anyone else’s timeline.
What OpenCode has not published
There is no post-mortem. As of 21 October, OpenCode’s changelog runs to v2.1.2 with no incident entry, and the company’s own post says only that “the incident is ongoing”, adding that it is “working with Anthropic, OpenAI, and Google to remediate”.
TechCrunch reported that OpenCode has not detailed how attackers gained access to the system, how keys were stored, or the full extent of any misuse. The facts on the record are fully explained by ordinary credential theft: someone read stored keys and spent them. Nothing public connects the stolen keys to model training.
OpenCode called the breach “a large scale distillation attack”. Nathan Lambert, an open-model researcher at Trillium Labs who writes Interconnects, pushed back on 17 October: “There’s something very odd going on with ‘distillation’ if a cyber attack to steal api keys is looped under the same term… It’s not clear why the volume is so high and how it’s related to model training.” TechCrunch reported that sceptics called the framing an overstretch and that any use of the keys for distillation remains unproven.
One widely repeated detail is absent from the primary record. Secondary coverage describes the breached system as a legacy console being decommissioned; TechCrunch describes it only as an older system, and OpenCode’s post does not describe it at all. OpenCode’s v2.1.2 changelog does mention fixing legacy console.opencode.ai credentials not being forwarded to the new console, which corroborates that a migration happened. Aggregator coverage, including Notebookcheck, The Register and SecurityAffairs, recycles the same primary material and adds no new evidence.
The guardrail claim and what it proves
Asked whether frontier models were blocking cyber work, Raad replied on 18 October: “yes that was a bit frustrating mid incident when trying to investigate. they did quite a bit and then started refusing”, adding that the team had gained access to Daybreak Blue and that things were “better”. That reply is the entire basis for the claim. Raad did not say why the models refused and never mentioned safety filters. The mechanism, that the malicious code the team asked models to examine tripped safety filters, comes from TechCrunch’s reporting of the company’s account. The defensible version is that OpenCode says its own investigation was interrupted by model refusals, and that TechCrunch reports the company attributes that to safety filters.
Daybreak Blue does not settle the claim. It is a tier of OpenAI’s Daybreak programme for approved defensive cybersecurity work, offering flagship models with reduced refusals for authorised workflows such as malware analysis in a controlled environment and incident response. Access is approved and provisioned per identity, organisation and project, and OpenAI’s documentation states that a selected programme “doesn’t guarantee that every safety check or prompt will succeed”. As of 21 October, the Blue model alias is marked deprecated.
A real industry pattern sits under the anecdote, though it is a pattern rather than proof about this incident. Hugging Face’s first-party post-mortem from July 2026 describes its own log analysis being blocked by providers’ safety guardrails that, in Hugging Face’s words, “cannot distinguish an incident responder from an attacker”. Hugging Face ran the forensics on an open-weight model on its own infrastructure instead, and argued that defenders should have a capable model they can run themselves before an incident. TechCrunch documented the same complaint from named researchers in July 2026. The two events are separate: Hugging Face was dealing with an intrusion by autonomous agents, OpenCode with stolen credentials from a hosted console.
David Sacks, co-chair of the President’s Council of Advisors on Science and Technology, amplified the story with “The guardrails backfired.” That is commentary on two incidents, not evidence about OpenCode’s.
Where the key lives is a design decision
OpenCode’s homepage says the product “does not store any of your code or context data”. That promise covers code and context, while the Console documentation separately discloses that Console stores provider credentials and serves each provider through its gateway. Different objects, though the headline promise reads broader than the credential design behind it.
A provider key you keep on your own machine is exposed if your machine is compromised. A key you hand to a hosted console is exposed if the console is compromised. Console is optional, and the app and CLI do not require it.
The poisoned LiteLLM release shows another way coding-tool credentials can leak, through software installed in the developer’s environment.
