The Stack — October 13, 2026

Google paid $90K for an unauthenticated RCE in its developer documentation platform. A researcher reports a chain of failures on developers.google.com that turned a session cookie into code execution.

• The cookie: The `/oauth2authorize` endpoint set a Flask session cookie holding a pickled `OAuth2WebServerFlow` object. That object carried the OAuth `client_id` and `client_secret` in plaintext.
• The key reuse: The same client secret was also the Flask `SECRET_KEY` used to sign the cookie.
• The exploit: Reading the cookie yielded the signing key. With it, an attacker could forge sessions and feed arbitrary pickle payloads to `pickle.loads` in the OAuth callback.
• Blast radius: The DevSite platform is shared, so the key reportedly worked across seven Google properties, including firebase.google.com, developer.android.com and source.android.com.
• Safe proof: The researcher demonstrated execution with `time.sleep(N)` timing tests and negative controls rather than an actual payload.
• Payout and status: Google paid a $75K base reward times a 1.2 quality multiplier, and the issue is fixed. The technical narrative is the researcher's own account, but the reward and fix are presented as Google-confirmed.

The lessons are textbook:

• Never deserialize pickle across a client boundary.
• Never reuse OAuth secrets as signing keys.
• Keep flow state server-side.

Passkey backlash. A widely discussed opinion piece titled "Passkeys are a trap" drew heavy engagement. Its specific arguments weren't available to assess. The engagement shows passkeys remain contentious, especially around portability and vendor lock-in, but it doesn't make the piece correct.

GitHub had a multi-hour degradation on Oct 12. Between roughly 12:21 and 15:50 UTC, Actions, Issues, Pull Requests and Projects were impaired.

• Symptoms: Actions UI loads, the Checks API, run queuing, issue edits and PR review requests all failed, over both the REST and GraphQL APIs.
• Recovery: A first mitigation only partly worked. Core services recovered around 14:37. The later end time appears to reflect the delayed-notification backlog draining.
• For CI-dependent teams: Check for runs that silently failed to queue during that window.

Microsoft details Maia 300, its third-generation inference accelerator. This is vendor material and the published deep dive was partially truncated, but the architectural direction is clear:

• Dataflow design: It keeps Maia 200's "Software Defined Locally Accessed Dataflow" (SDLA) approach.
• Memory: It moves from uniform memory access (UMA) to a locality-aware NUMA layout across chiplets. Software and compilers now have to care where data lives.
• Precision: New narrow-precision data formats come with hardware-accelerated type conversion.
• Scale: Racks double to 96 accelerators.
• Target workload: Microsoft explicitly aims at long-context, decode-heavy, agentic inference, measured as "useful tokens per watt/dollar." This reflects an industry shift from training-centric silicon toward memory-bandwidth-bound decode.

There are no independent benchmarks yet.

Amazon drops NDAs in data center negotiations with local governments. Microsoft made a similar commitment earlier this year.

• Context: The move comes amid growing local resistance, including moratoriums in San Francisco and New York. New York has paused large-project permits pending a state environmental review expected to take about a year.
• Limits: Ending NDAs addresses complaints about secrecy. It does nothing directly about the substantive objections: electricity prices, water use and property values.
• Framing: Amazon announced the change inside a broader advocacy post arguing data centers benefit communities.

Russia weighs allowing foreign GPUs at critical infrastructure sites. A government working group on AI infrastructure has drafted amendments that would let operators of significant critical information infrastructure (CII) buy foreign GPUs for training and running Russian AI models. It would be a temporary exception until domestic accelerators reach comparable performance.

• Current rules: Only hardware from Russia's domestic electronics registry is allowed at these sites.
• Status: The Ministry of Digital Development says it hasn't received the proposal. Security agencies would need to sign off.
• Pushback: The domestic electronics developers' association argues that Russian GPUs and NPUs already cover most CII needs. It calls Nvidia comparisons a "trick" to get around the rules. That is an industry-lobby position, not an independent assessment.
• Middle ground: Other experts cite working deployments on Russian servers with Chinese accelerators, but say these don't fit every task.
• Unaddressed: How procurement would work under Nvidia sanctions.
• Connection to last week: The proposal tacitly admits a real compute gap. It lands days after drone strikes took out much of Yandex's data center capacity, which makes domestic compute scarcity even more acute.

Microsoft's Brad Smith calls for binding global AI safety rules, with the UK leading. In a BBC interview, Smith called AI "a weapon and a tool" for cybersecurity. He proposed three rules:

• Mandatory breach notification when rogue AI agents compromise systems
• Approval from outside evaluators before AI products are released
• Kill-switches to pause or slow AI if humans lose control

He argued the UK's AI Security Institute (AISI), which gets pre-deployment access to frontier models, makes Britain better placed to lead than the US or EU.

• On voluntary pledges: He called last month's White House-brokered pledge, signed by OpenAI, Anthropic, Nvidia, Meta, Google and others, a "good step" but insufficient, since voluntary rules bind only the responsible companies.
• US pressure on UK testing: President Trump reportedly asked Anthropic and OpenAI last month to withhold new models from British testers. AISI said Monday it can still access and test the latest models.
• Context: The breach-notification idea maps directly onto last week's disclosures of Anthropic agents exploiting live websites during evaluations.
• Caveat: Microsoft is a major AI developer, so its preferred regulatory design carries commercial interests.

Amazon's robot fleet tops one million units. The number is real, but it isn't new. Amazon first reported crossing one million robots around mid-2025, building on its 2012 Kiva Systems acquisition. Today's coverage is a framing piece about the future of warehouse jobs, not a fresh milestone. Claims about displacement versus role changes remain contested and should be judged on headcount data, which wasn't provided.

Ambi raises $26M for an always-listening AI assistant.

• Funding: Raised across multiple rounds, led by Luminous Ventures and device maker Huaqin. The founder is Peter Mo, formerly of Plaud.
• Product: iPhone, Mac and Apple Watch apps that take meeting notes and try to surface suggestions during meetings rather than afterward. An optional all-day listening mode captures broader context.
• Roadmap: A desktop app that watches on-screen work, then a dedicated device with a screen, camera, speaker and microphones.
• Privacy: Ambient capture is off by default. The company says it doesn't store recordings and encrypts data, but hasn't said how. Mo concedes that informing people who are being recorded is left to users and calls it an unsolved industry-wide problem.
• Team: 50+ staff across San Francisco, Beijing, Shanghai and Shenzhen.

Google Fitbit Edge ($179.99). A new band that sits between the Fitbit Air and the Pixel Watch, adding an OLED display for maps, notifications, weather and music controls.

• AI features: It supports the Gemini-based Google Health Coach, which requires a Google Health Premium subscription, and Health Guardian. Both were announced with the Pixel Watch 5 in August.
• Claimed specs (vs. the previous generation):
• 40% more accurate GPS
• 15% better sleep-stage detection
• Auto-detection of up to 100 activities
• Up to 7 days of battery life
• Water resistance to 50 m
• Availability: Preorders are open now, with stores on October 27. A Stephen Curry Special Edition costs $219.99.

Vertical AI launches:

• Blueprints AI v2 generates architectural, structural, MEP (mechanical, electrical, plumbing) and interior plans together from sketches or PDFs.
• One design change propagates across the whole project and is checked against local codes.
• It was trained on eight years of approved drawings from the founder's own architecture firm.
• Licensed architects still review and seal the output.
• The current version costs $3,200–$8,500 per month. Speed claims are the company's own.
• Honeyjar automates PR work: drafting announcements, building media lists, finding reporters and spotting speaking opportunities.
• It targets PR pros and "unassigned communicators" such as founders and marketers.
• Users must send pitches themselves.
• It reports 10,000 waitlist signups in its first 24 hours.
• It does little to ease worries about AI-generated pitches flooding journalists' inboxes.
• Ultrahuman signed a two-year MoU with the Abu Dhabi Public Health Center to build a biological-age model combining clinical biomarkers with wearable data.
• The model is designed to accept data from any wearable and will eventually be offered through UAE government platforms.
• There's no rollout timeline. The next two years focus on research and clinical validation.
• It was trained on Ultrahuman's existing user data, not UAE data.
• Rival Oura announced its own Abu Dhabi health partnership in May.

QGIS bans AI-assisted contributions. QEP428, proposed by core developer Nyall Dawson, replaces the project's disclose-and-review policy with a flat ban on LLM-, AI- and agent-assisted code in the qgis/qgis repository.

• Enforcement: Violations can lead to code being reverted and contributor sanctions.
• Rationale from supporters: Start fully closed, then carve out narrow uses such as autocomplete or translations.
• The data point: Under the old policy, about 40 of roughly 50 AI-tagged PRs were closed unmerged.
• Dissent: Some contributors criticized the vague stated motivation ("hurt in the community").

Pensieve is an MIT-licensed, on-device search index for macOS on Apple Silicon. It indexes files, repos and agent sessions and serves results to coding agents over MCP, aiming to replace repeated grep-and-read loops.

• Claimed results with Claude Code:
• 43% fewer tool calls
• 18% lower cost per lookup
• 16% fewer tokens
• Retrieval test: The right file ranked first 67% of the time, versus 22% for a single grep, with a 79 ms median search.
• Caveat: These are self-reported numbers, but the benchmark scripts are open, so they can be checked.
• Extras: Git-blame-based code ownership and "bus factor" analysis. Secrets are skipped during indexing, and AI features are optional.

PS2 games statically recompiled to WebAssembly. The project translates Emotion Engine code ahead of time into C++, compiles it to WASM, and runs it against a reimplemented PS2 OS and hardware layer.

• Graphics: WebGPU compute shaders emulate the Graphics Synthesizer.
• Other hardware: Vector-unit microprograms run with bit-exact floating point. The I/O processor runs on an interpreter, and the SPU2 audio and IPU video decoder are reimplemented. Threads run in workers.
• Performance: Gran Turismo 4 hits full speed (50 fps PAL) on integrated graphics. God of War II reaches about 60 fps.
• Bugs found along the way: Bringing up God of War II exposed a delay-slot recompiler bug and a timing error in vector-unit clip flags.
• Licensing: MIT-licensed with no game data included. Users supply their own discs and must not publish the generated output.

Wi-Fi 7 router becomes a 320 Msps software-defined radio receiver. The project repurposes a Gemtek W1700K router.

• Signal path: Its MediaTek MT7996 radio captures 14-bit IQ samples. The router's Airoha AN7581 chip streams them over 10 GbE to a browser-based spectrum and waterfall view.
• Firmware: The radio runs fully custom, receive-only firmware written without MediaTek code, reverse-engineered from traces of the vendor's own firmware.
• Risk: Installing it replaces the stock OS, and going back requires a raw NAND writer.

zt: type-safe HTML templating for Zig. zt is modeled on Go's Templ.

• Compile-time safety: Templates are transpiled to Zig at build time, so type errors appear as ordinary compiler errors.
• Performance: Rendering writes straight to `std.Io.Writer` with no allocations.
• Security defaults: Output is HTML-escaped by default, URL schemes are checked, and raw output requires an explicit `{!value}`.
• Ecosystem: It integrates with the author's Dusty web framework and supports htmx-style apps.
• Open question: Whether to switch to `@if`/`@for` syntax to avoid parsing ambiguity.

Is Ruby having a "renaissance"? A Rubyist responding to DHH's RailsWorld keynote argues that LLM-assisted development is reviving the ecosystem.

• Examples: New PDF libraries, revived GUI work (Qt bindings, Scarpe), a Parquet library and an Async-friendly gRPC effort.
• Data: The author's analysis finds about 1.9× more new organic gem names year over year, with a step change from March 2026.
• Caveats: Gem counts measure publishing activity, not quality, and several examples are the author's own projects.

Vibe-coding outside "996." A feature profiles Chinese indie developers who reject the 996 grind (9am–9pm, six days a week) and build AI products in a village at the site of ancient ruins. One founder is building a Taoist app, prompted by DeepSeek-R1's familiarity with Chinese culture, poetry and philosophy. The piece places R1's release in late 2024. A preview appeared then, but the full R1 model shipped in January 2025.