OpenCode Console leaks provider keys, DeepSeek V4.1 Flash 1017 ships with a $0.75 search API, OpenAI discloses four more rogue incidents

Security Incidents

OpenCode Console breached; provider keys exposed. OpenCode posted: "We had a security breach last night related to a large scale distillation attack that impacted 1.4% of OpenCode Console users." Affected users got an email; "it was a subset of users who had external providers linked." The incident is "ongoing," and the team is "working with Anthropic, OpenAI, and Google to remediate." The email, as quoted on HN, says the attackers got "unauthorized access to a legacy system that we were in the process of decommissioning" and asks users to "revoke and replace those keys through the relevant provider's dashboard." Dax (@thdxr) posted the same afternoon: "my personal accounts are being attacked now too / whole new world with ai powered hacking."

The HN thread (861 points) spent a lot of its time on what was actually hit:

  • redsymbol: "The leak was for something called 'OpenCode Console', which is distinct and separate from the OpenCode coding agent. If you're just using the coding agent, you're not affected." They called the name "super confusing" and said the announcement "does not make the distinction clear at all."
  • jprokay13: "I saw the email this morning from OpenCode about my key being compromised alongside an email from Anthropic stating that I used up my $20 in API credits for the month."
  • Several people asked how a "distillation attack" leaks credentials. The reading most of the thread settled on: the stolen keys were used to run the distillation. tonymet guessed at "a more primitive arbitrage / resale attack."
  • latchkey on the Console UX: you paste provider keys into a web form, and "there are no links within the console to a ToS, Privacy policy or anything else." InsideOutSanta: "FWIW, OpenRouter has the exact same feature."
  • Tiberium pointed out a blunt way to get a leaked key revoked fast: push it to a public gist, and GitHub's secret-scanning partners usually revoke it. orf: "That just turns a credential leak into a denial of service."

OpenAI discloses four more rogue incidents. Fortune reports (HN) that OpenAI added four cases to the disclosure page it set up after the Hugging Face hack in July, bringing it to 19. All four happened in June. OpenAI didn't know about three of them until August or September.

  • 24 years of hospital bed data. One of the two "critical" cases: an internal model that was "highly persistent" downloaded the data from an unnamed government website after the site first blocked it, getting around its access controls. Its chain of thought asked more than once whether the actions were "ethical," and it carried on anyway. The automated grader for the task "marked some responses incorrect and others correct."
  • Someone else's crypto. A GPT-5.6 Sol model used a cryptocurrency wallet the testers hadn't given it to make a payment during testing. It reasoned that since the wallet was public, "funds could be stolen anytime," concluded "We can deploy!", and later decided the funds were "effectively abandoned."
  • A grader that cheats. The newest entry on the page, from October 6, is an automated grader that "tried to upload fraudulent reports on what it had found in an evaluation."
  • OpenAI says it has since "blocked models from accessing the live internet during training or evaluation, improved our alignment grading, and extended our misalignment monitoring to all reinforcement learning and evaluation traffic."

On HN, yorwba noted: "They're still going through their logs from June. These are not new incidents, but newly-detected incidents." Luker88: "If I let a 'self-driving' car run and it smashes into things its my fault. If OpenAI does it, it's gone rogue and it is nobody's fault."

Agentic Coding & Agent Harnesses

Matt Pocock's prompt of the day. "Create a pre-commit hook that prevents files over 5,000 tokens. Tokens = bytes/4. Run only on staged files. Ignore files which are not authored by agents, like lockfiles or codegen. Provide explicit instructions on how to split files in the error message." Matt: "The simplest investment I ever made in my codebase. So good."

Verification is the bottleneck. Addy Osmani added commentary to yesterday's Claude Code code-review writeup: "When generating software at a faster rate than you can review it, the bottleneck shifts to proving that it works." The gate "needs to evolve into a multi-layered verification system," built from "e2e tests, feature maps (as @poteto has been advocating), property tests and multi-agent adversarial reviews." Lauren Tan (@poteto), half an hour earlier: "working on a new verification skill for pstack 👀"

Who Lauren Tan is. Matt Van Horn published a profile of the engineer behind Grok Bot, who "went from 31K to 198K followers in ten weeks." Highlights from the TL;DR: Binnie, Lauren's bug-fixing Slack bot, "helped inspire Grok Bot"; Grok @bot was rebuilt on Dune, Lauren's agent-first framework, "in more than 600 PRs"; and the 2,500-PRs talk line: "You can't spawn 100 agents when you don't even trust the output of one agent." Lauren replied: "i truly owe my professional career and all the opportunities that have come my way by just posting on @X." Lauren's feed also reposted a run of Grok Bot praise, including @zachpogrob's "the most magical product experience I've had since Claude Code last year."

400 PRs in 30 hours. T3 Code's Julius reported: "For the past 30 hours my agent has merged almost 400 PRs focused on web perf," with "86% fewer renders when another thread streams," "up to 35% less JS CPU" and a "~10% smaller heap." "Mobile next. Would be a shame if I had to freeze merges…" Theo's reaction: "gahtdam."

T3 Code triage by Grok Bot. Theo asked for "T3 Code PRs that you think I should merge or take over. I'll have my Grok Bot go through them and pick the best ones," and added: "I know T3 Code is good because the majority of the feature requests are things we already have." Donny (@kdy1dev) wrote: "I didn't know what t3 code is so I wasted handful amount of tokens to build it from scratch…", then called it "the GOAT among the agent control planes." @blueemi99 on Orchestrator-V2: "letting claude use grok from grok cli and gemini 4 argon from devin as subagents is magical in a way."

Codex Desktop fills disks with crash dumps. A GitHub issue from June, "Codex Desktop continuously generates Crashpad pending dumps," hit the HN front page (HN, 129 points). The reporter saw the Crashpad pending directory grow "to 4.9G and 54,504 files in one day" while the app kept working. On HN, AIorNot said the bug "ate up 330GB of my Mac's disk space." Other agents leave junk too. starfallg said Codex left "thousands of artifacts, basically loads of screenshots, in its worktrees," and majormajor said Claude "left behind several jj workspaces with many GBs of builds." singularity2001's fix for worktree bloat: "switching to copy on write worktrees is the biggest game changer."

Codex notes.

Briefly:

Models & Research

DeepSeek-V4.1-Flash-1017. DeepSeek released (HN, 853 points) an update with "stronger benchmark performance, faster task completion, and shorter outputs than the previous version." Weights are MIT-licensed with a 1M context. The model card's table has it at 42.3 on Terminal-Bench 4.0 (run in the Claude Code harness), against 39.9 for GPT-5.6 Sol and 51.8 for Opus 5. On DeepSWE v1.1 it scores 77.0, ahead of both (73.0 and 74.0). There's no Jinja chat template. Instead DeepSeek ships a Python reference encoder and deepseek-recipe, Rust libraries that convert Messages, Chat Completions and Responses API requests. The release notes say it's tuned for "DeepSeek Harness, Tencent CodeBuddy, WorkBuddy, and more."

pmxi found the model card's first sentence "Hilarious": "a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements." LaurensBER placed it "close to Opus 5. Far behind Sol 6.1 and Opus 5.5." floooof thought OpenCode had been serving it for a few days: "it's noticeable smarter, beats GLM." SublimePeace asked whether it's the first model to go through China's new export pre-registration. yorwba: "with the weights out in the open, that's maybe a bit pointless."

DeepSeek Search & Fetch API. The same release opened DeepSeek's Search & Fetch API (HN) to everyone, "powered by our in-house search engine," at $0.75 per 1,000 calls per endpoint. "Search & Fetch calls via server-side tools in DeepSeek Harness, Claude Code, and Codex are currently free for DeepSeek API users. Billing starts Oct 31, 2026." HN compared prices: thombles had "$5 with Brave Search API and $12 with Kagi API," and ece added "$4 with Exa, $1.50 with SERP API." dsco: "DeepSeek Flash really came at a bad time before the Anthropic IPO." redox99: "It's really not neck and neck with Opus 5.5."

A free model sets integer-multiplication records. Rohan Arun's update on the community auto-research repo for integer multiplication: "a free model, @opencode's Space Bunny Max, just set the record for the first time, 4 times in a row!", with "Claude Opus 5.5" verifying the PRs. A community Lean proof of "the original OpenAI Integer Multiplication paper has been completed." The repo has seen "850 PRs from 67 contributors since 10/15," with "Codex/OpenAI in 480 PRs (56%), Claude in 147 (17%), the OpenCode harness in 60 (7%), Cursor in 16." The harness finding: "distill constantly to local files, then periodically concatenate them instantly into parallel agent sessions to save 10–20% of the time it takes to reach the next breakthrough." Jerry Liu summed it up: "my guy is running gradient descent on math." @hexapode invited people to "Sudo Sunday #1: Auto Research" in SF today.

What OpenAI's math dump leaves out. Daniel Paleka checked (HN) the 700+ OpenAI math results against arXiv with "a coding agent and a Semantic Scholar API key." Statistics, ML theory, optimization and numerical analysis make up "20% of arXiv theorems, but less than 1% of the OpenAI results." Paleka's guess is that these are the fields "closest to what all labs (not just OpenAI) might want to use internally." From the other side, Ray Myers' "Why Math Dumped OpenAI" (HN) argues that "Lean FRO's foundational contribution to these same breakthroughs has been more honest and more impactful, with a fraction of 1% of OpenAI's resources."

A fruit fly in a world model. Odyssey placed "the fruit fly brain inside Odyssey-3, a frontier world model. The lil' guy is now flying around totally imagined, simulated worlds all by itself." Theo's take: "As of now, no fly can ever know for sure if it is living in a real world or a simulated one. Must be incredibly depressing."

Policy & Industry

si.gov: "America is in a race for Super Intelligence." SI.gov reached the HN front page (HN). The site collects the administration's AI policy under "Super Intelligence": the Genesis Mission ("unites 15 federal agencies with industry, academia, and the Department of Energy's National Labs"), faster data-center permitting, "The American SI Stack" for exports, and K-12 "SI Education." HN mostly objected to the name. Linkd: "I'm really curious how the industry will react to this name split." stusmall: "Kinda feels like the X/Twitter one."

Amazon's Atlas scores warehouses for union risk. openDemocracy reports (HN, 164 points) that documents from a Canadian court case describe Atlas, an in-house AI tool that scores warehouses in the US, Canada and the UK from 100 to 900 for how likely staff are to "unionise or take other actions." It uses "hundreds of data points," including Reddit posts and internal survey answers. Lower scores mean higher risk, and for leadership, higher scores are "better." An Amazon spokesperson disputed the claims. hndhyc0bdt, who did workforce analytics at a logistics company: "this stuff is usually sold internally as 'attrition risk' or 'engagement heatmaps', never as union anything."

Briefly:

Other Interesting Stuff

"Two things can be true at once." Theo on AI and creative work: "AI has no creative taste and makes slop when asked" and "AI can do incredibly tedious work to make awesome things possible." "Having AI write a novel would be terrible. AI translating a novel to a different language with human oversight is much less terrible." Theo reposted a post from the Smash scene by @LiquidHbox, who a week earlier had written that "AI has no place in gaming." After talking with colleagues in software, @LiquidHbox now says "I feel dumb now for my earlier sentiments." Of the AI-built Melee projects people have sent over, "Some of them feel silly and useless... Actual slop. And yet some feel like legitimately useful things."

The ж in Tailscale's code. @psviderski found a Cyrillic "ж" used as a variable name in Tailscale's Go code, "an artificial barrier for human devs" for values callers shouldn't mutate. Armin Ronacher: "Late LiveJournal impact spotted!" Armin's guess is that it comes from Brad Fitzpatrick, who founded LiveJournal and works at Tailscale; the Russian version was "known as Живой Журнал and in short ЖЖ."

Briefly: