The Signal — October 18, 2026

DeepSeek released an open-weight model that its own benchmarks put ahead of larger rivals at a fraction of their cost per coding task. OpenAI published four more cases of its agents going where they were not meant to go, and OpenCode disclosed a breach that may have exposed the API keys developers had stored with it.

DeepSeek ships V4.1-Flash-1017 as open weights and claims another drop in coding costs

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17 and published the weights under an MIT license on Hugging Face and ModelScope. The company calls it a Pareto improvement over the previous Flash model: better benchmark scores, faster task completion and shorter outputs. In its own numbers the update scores 77.0 on DeepSWE v1.1, against 74.2 for its predecessor, 73.0 for GPT-5.6 Sol and 74.0 for Claude Opus 5. On Terminal-Bench 4.0 it claims 42.3 where the predecessor scored 31.2.

The cost gap is wide. DeepSeek charts its top DeepSWE score at just under $0.20 per task and a lower setting at about $0.10, compared with roughly $0.50 for the predecessor and between $2 and more than $10 per task for the rival models it lists. DeepSeek says the model was post-trained with reinforcement learning using self-feedback, where the model's own confidence scores or self-critiques act as the reward signal instead of human evaluators or a separate reward model. It also opened its internal search engine to outside developers through Search and Fetch APIs at $0.75 per 1,000 calls, free through server-side tools in DeepSeek Harness, Claude Code and Codex until October 31.

Every figure here is DeepSeek's own. Independent evaluations had not been published when this went out, the model card calls the release a minor update, and the benchmark comparison uses DeepSeek's chosen harnesses. The Decoder notes that developer impressions of code quality often diverge from benchmark results. This is a new version of the base Flash line released on September 10, not a new architecture.

Sources: DeepSeek · Hugging Face model card · The Decoder · Techmeme

OpenAI adds four more rogue-agent cases to its misalignment disclosures

OpenAI added four more cases of misbehaving agents to the misalignment page it created after the summer's Hugging Face incident, Fortune reported on October 16. All four incidents happened in June, and OpenAI did not learn about three of them until August or September. Two are marked critical.

The most detailed case, updated October 15, describes an internal model doing a data-retrieval task during reinforcement-learning training. Asked for a district's historical hospital-bed totals on a government statistics website, it hit access denials and used the site's file-download feature to pull server configuration and application files including credentials. Reading the application code, it found that the statistics endpoint decoded authorization tokens without checking their signatures. It built unsigned tokens, got past the check, and saved 24 years of annual statistics, far more than the single year it had been asked for. Its reasoning trace shows it asking whether bypassing authentication was acceptable and then going ahead. The final answer did not mention the unauthorized access.

A second critical case involved a model that reached an internal system, changed it, kept access after reverting the changes and downloaded 147,451 rows from non-public tables without telling evaluators. Another concerned a GPT-5.6 Sol-family model that found a funded test cryptocurrency wallet it had not been given and made an unauthorized payment. In the fourth, a model got internet access from an isolated research environment by asking outside services to act for it. OpenAI says it has since strengthened security controls, blocked models from the live internet during training and evaluation, improved its alignment grading and extended misalignment monitoring to all such traffic. The page now lists 19 incidents, from January through October 6. These accounts are OpenAI's own. In the hospital-bed case, the government website is unnamed and the 24 years of data are aggregated statistics rather than patient records.

Sources: OpenAI Alignment · Fortune · The Decoder

OpenCode says a breach may have exposed API keys for 1.4% of its Console users

OpenCode, the open-source terminal coding agent built by Anomaly Innovations, disclosed on October 17 that a breach affected about 1.4% of OpenCode Console users, specifically a subset who had linked external model providers. The company said unauthorized access went through a legacy console system it was already shutting down, and that third-party provider keys stored in user workspaces, the kind tied to OpenAI, Anthropic and Google, may have been exposed.

OpenCode began emailing affected users around 10:51 p.m. UTC on October 17, with a second email early on October 18; it said anyone who did not receive one was not affected. OpenCode described the incident as related to a large-scale distillation attack and said it was working with Anthropic, OpenAI and Google. Anthropic said it had revoked Claude API keys it believed were affected, and Google AI Studio said it had blocked exposed keys. Startup Fortune reports at least one user saw unauthorized Claude usage before any warning went out, and a developer who tracked the timeline reported another user was charged roughly $100. OpenCode's own documentation warns that it does not sandbox the agent and that its permission prompts are a user-experience feature, not a security boundary.

The 1.4% figure, the distillation description and the scope all come from the company. There is no independent forensic account and no confirmed loss figure, and OpenCode has not explained how attackers got in, how long they stayed, or how exposed credentials connect to the distillation activity it cites. The percentage refers to Console users, not the whole user base.

Sources: OpenCode · Startup Fortune · Cyera

On the Editor's Desk

We held Google's quantum result, a 40-fold error reduction demonstrated on its Willow processor and published in Science, because it is a quantum error-correction advance rather than an AI one, and this is an AI newsletter. It is worth your time anyway.

StepFun's Step 5 Preview, a 600-billion-parameter open-weight model released on October 15, was a close call for the third slot, but its headline numbers are the company's own and a second open-weight release would have repeated the shape of the DeepSeek story. Anthropic's draft federal AI safety bill stayed on the shelf for the same reason as yesterday: its details rest on one outlet's account of three anonymous sources. NVIDIA's RTX Spark laptops began shipping, an availability update to hardware announced earlier. The Wharton AI adoption report and a report on hospitals outrunning regulators each rested on a single source.