Quick Hits

• @opencode confirmed a "large scale distillation attack" affecting 1.4% of OpenCode Console users, specifically those who had linked external providers. @sudoingX relays user reports of $400 and $1,700 in overnight API charges and argues for keeping keys on your own machine.
• @pwn_ai disclosed PwnEscape, a KVM guest-to-host escape via a Linux kernel networking flaw (CVE-2026-99999) that reaches root execution on hosts. @PaulosYibelo expects VPS and compute engineers to lose their weekend to hypervisor reboots.
• @josesilesdata reports AnyPS5 runs PS5 games natively on Windows and Linux by translating executables instead of emulating the console, with games already hitting 60 fps and 27,800 GitHub stars.
• Two @TimSweeneyEpic posts circulated widely: one arguing AI should absorb the routine 70% of engine coding, another predicting "a new golden age of gaming" as previously uneconomical games become viable.
• @TeksEdge says Mistral shipped a 209M-parameter time-series foundation model that forecasts in roughly 140 ms on CPU, claiming #1 zero-shot results on three public benchmarks.

A Key-Draining Breach and a VM Escape Made It a Patching Weekend

The details from @opencode are thin because the incident is ongoing: affected users were notified by email, the impact was limited to a subset of Console users with external providers linked, and the team says it is working with Anthropic, OpenAI, and Google to remediate. @sudoingX fills in the texture: one user says keys used only in opencode ran up over $400 overnight, another reported $1,700, and every affected account had external providers linked. Worth noting the tension in framing: a "distillation attack" usually describes extracting model behavior, but the reported damage here is drained API keys. Until opencode publishes more, treat both the term and the numbers as preliminary.

The same weekend, @pwn_ai published PwnEscape, describing a path from untrusted VM code to host-kernel memory reads, controlled writes, and root command execution via a Linux kernel networking flaw affecting "industry standard hypervisors." @PaulosYibelo's reaction is sympathy for the admins doing emergency patching and rebooting, noting security bugs have a habit of dropping right before weekends. Two incidents, both touching the infrastructure layer, is a rough couple of days for anyone running compute.

AnyPS5 Translates PS5 Games Into Native PC Programs

Per @josesilesdata, AnyPS5 skips emulation entirely: a "relinker" converts the PS5 executable to your system's native format, the console's internal libraries get rewritten, shaders are recompiled to Vulkan, and input works through SDL (sticks and triggers) or keyboard and mouse. The author likens it to the approach that let Windows games run on Linux. The reported traction: games running at 60 fps, a compatibility list growing weekly, nearly 5,000 commits, 27,800 stars, 2,400 forks, GPL-2.0 license. All of those numbers are the author's, with no independent verification in this feed, but the mechanism described is coherent.

The same day brought the inversion: @0x0SojalSec posted a PS5 booting Windows 11, launching Ryujinx to play Hollow Knight, then casually returning to the PS5 dashboard ("either peak tech or the start of the simulation collapsing"). The same account also claims inFAMOUS Second Son now runs on Windows and Linux through AI-assisted reverse engineering. PS5 titles becoming PC programs while a PS5 becomes a PC, in the same feed, is a fitting snapshot of where compatibility work is headed.

Tim Sweeney Wants AI to Take the 70 Percent of Coding That Was Never Hard

In a long post quoted by both @petergyang and @boneGPT, @TimSweeneyEpic recalls writing Unreal Engine 1: roughly 30% of his time went to genuinely hard problems (assembly texture mappers, C++ BSP trees, networking), while 70% was typing and refactoring large bodies of code, which he says "didn't feel much different from vibe coding today." He argues it was always logical for AI to take that 70%, and that the real surprise is AI acting as an "insta-Carmack" that solves hard technical problems cleanly and works in complex 3D scenes, both by operating tools and by reading raw data files. He closes with an analogy to abandoning assembly in 1998: what matters is the end product.

A second @TimSweeneyEpic post, quoted by @dhh, predicts the games that never got built because they would cost several times more than they'd earn will soon become possible. @dhh compresses it to "AGE_OF_AGENTS == AGE_OF_ABUNDANCE." @petergyang hopes the StarCraft 3 team is reading and ships before 2030, while @boneGPT adds the origin story, from Jazz Jackrabbit and Jill of the Jungle to Fortnite.

Agent Tooling Pushes Into Daily Enterprise Work

Several posts point the same direction: agents moving from demos into routine work. @bcherny quotes @ClaudeDevs announcing that Code Review is generally available for Claude Team and Enterprise plans, with faster reviews and admin-selectable Quick, Standard, or Deep levels per repo; @bcherny says it runs on all his PRs and "catches almost every bug," calling it part of his team's secret sauce.

On the Microsoft side, @davidfowl quotes @DanWahlin on Work IQ in the GitHub Copilot app: Wahlin batch-created GitHub issues linked to M365 files in under 30 minutes instead of hours, including a reusable skill for future syncs, while @davidfowl calls the broader feature set (meeting booking, scanning Teams chats and meetings from an agent CLI) a game changer for the Microsoft enterprise ecosystem.

Scale is improving too. @BrendanFalk quotes @mxstbr confirming ChatGPT and Codex now use code mode and tool search, that the ~2k tools-per-MCP-server limit mainly guards against abuse, and that Cloudflare is about to serve its raw toolset of 3k tools. Falk's takeaway: models handling 3k exposed tools just fine is the notable part.

Two posts on the people behind the tools: @mvanhorn profiles Lauren Tan (@poteto), the Cursor engineer behind Grok Bot, covering her path from an Excel macro to the React Compiler, rebuilding Grok Bot on her Dune framework across 600+ PRs, and her own caveat that she didn't build it alone. And @distortgeekin packages an unverifiable anecdote (a "SpaceXAI engineer" running 200+ agents that audit each other nightly) into a pasteable prompt; treat the anecdote as marketing, but the underlying loop, turning mistakes into persistent rules, is the reusable idea.

Briefs: A Cancelled Booth, Reset Wars, Open Weights, and a Hot Plate

• @GergelyOrosz reads AWS's last-minute cancellation of Cloudflare's Re:Invent booth, which @eastdakota says Cloudflare will work around with free side events, as a sign AWS treats Cloudflare as a competitor, since cancelling a partner booth is "unheard of."
• @_mohansolo announced quota resets for Argon and Gemini on Antigrativity's weekend build push (note: the platform is Antigravity); @beffjezos quipped "Google has its own Tibo now. Let the reset wars begin."
• @finkd says Muse Spark Open weights release today and run great on new Mac Studios thanks to work with Apple, shared by @oprydai mostly for the "billionaire desk setup" joke.
• More Mistral detail from @TeksEdge: the 209M model is 837MB in FP32, ingests up to 32,768 historical time steps, forecasts up to 1,024 steps per forward pass, and outputs 21 probability quantiles; a 5M variant (Timestral-5M) shrinks to 21MB and ~10 ms on an 8-thread CPU, per Mistral's claims.
• @lennysan shares @lil_dill's Summit talk on avoiding "zombie UIs," with four ideas: have a point of view, encode your standards into the machine, don't confuse done with good, and unleash artistry. The premise: using AI only to make existing things faster misses the interesting part.
• @badlogicgames endorses the USB-C hot plate from @MehdiHacks's backpack-sized soldering kit (hot tweezers, hot air gun, iron, mini plate), while admitting he can't quit his desktop station.

Practical Takeaway

The strongest signal this weekend is key hygiene. If you linked external providers to OpenCode Console, rotate those keys now even if you didn't receive the incident email; @opencode says only 1.4% of Console users were hit, but a rotation costs minutes and a drained key reportedly cost one user $1,700 overnight. More broadly, @sudoingX's advice generalizes to any hosted agent harness: keep provider keys on machines you control and scope them per project. And if you run hypervisors for others, patch before Monday.