Overnight the security week turned into game shows and ambushes: a live bounty box pays real bitcoin to the first person who pops a deliberately vulnerable PHP server, and a researcher published a three-day audit of Tailscale with five finding classes and no vendor coordination. In the quieter corners, Dream Labs came out of stealth with $90M and a model that drives robots it has never seen, Apollo’s chief economist asked where the AI payoff is hiding, and a language designer shipped a systems language with no heap at all.

Today: the bitcoin bounty, the zero-shot robot brain, two chart packs arguing about AI’s real economy, the no-heap language, the Tailscale audit, a model that thinks 39% less, and Claude Code with a video-editing skill.

Show, don’t tell: a bounty, a demo grid, an audit and a benchmark table.

🪙 Steal My Bitcoin: the PHP2Shell bounty box is live

The counter on the homepage reads “0.00 hours without a successful hack” as we write this. The setup fits in a paragraph: there is a bitcoin key on this server, worth about $500, accessible to the PHP user account through the print-flag binary. The PHP version was, in the site’s own words, “selected from before the PHP2SHELL CVEs were released.” First person to get the wallet wins, and if you don’t need the money, the site nudges you toward GiveWell. All traffic is logged, and the operator warns you up front not to submit anything you want to keep private.

It is the first live-fire public test of No. 019’s PHP2Shell, the CVSS 9.2 pair that dominated this week’s patch cycles. The ecosystem around that bug keeps growing: an animated explainer made the rounds overnight, and a developer’s PoC dig titled “I want my Wednesday back” asks why mass exploitation has not started yet, pointing at the best public exploit (9bie/php2shell-bypass-aslr-exp) and the friction still in it. Reuters added context of a different kind: the PHP Foundation, steward of the language, is running on one full-time staff member.

The bounty box  ·  HN discussion  ·  The PoC dig

🤖 Dream-1: a robot brain that skips the training montage

In a test clip from Kuala Lumpur, an arm at Eastworld’s lab places an apple into a brown box. In Shenzhen, a BitRobot setup pulls a tissue from a box. In Paris, a Hugging Face robot moves a red cube into a green container. Dream Labs’ release post lists a dozen beta sites like these, each running the same model on its own hardware through the public API, with no per-robot training. The small print is worth reading too: the partner clips play at 8x speed, the in-house ones at 4x.

Dream-1 is what the company calls a World Action Model: connect a new bimanual robot with parallel grippers to the API and it starts performing tasks out of the box, no fine-tuning, no demonstrations, from $3/hour. The pitch rests on two curves crossing. Hardware is collapsing in price: a capable two-arm setup now runs about $2,000, seven semi-humanoid robots sit under $15k, and Hugging Face’s $399 Microduck sold over 10,000 units in four days. And generalization is getting cheap: their February DreamZero paper, which introduced World Action Models, showed 30 minutes of data from a new robot was enough to transfer task knowledge to it. If both curves hold, robotics’ chicken-and-egg problem (no data without deployed robots, no deployed robots without usefulness) finally breaks. The company is six months old.

The research post  ·  HN discussion

📉 The AI payoff question, in two chart packs

Torsten Sløk, chief economist at Apollo, opened his weekend note with a sentence you would not expect from a firm that lives off market optimism: despite massive investment and hype, AI has delivered zero financial payoff to corporate America outside of tech. His evidence is three flat lines: revenue per worker rising for the Mag 7 but flat for the S&P 493, profit margins the same, and no sign of AI in productivity growth. The same weekend, HN sent a month-old analysis by Joseph Politano to the front page: America has lost more than 200,000 jobs in media, film and the arts over four years, about 50,000 of them in the last year alone.

Read as a pair, they bracket the AI-economy argument. Sløk’s side: mainstream corporate America is starting to treat token spending as a cost center rather than a growth engine, and “the AI story will be decided in the S&P 493, not the Mag 7.” His kicker is quotable: AI may be a great calculator, but in business intelligence is mostly judgment, and AI does not understand accountability, trust, culture, unwritten rules or fairness. Politano’s side: the creative-class losses are real but causally tangled, with consolidation, offshoring and the streaming squeeze in the mix. His tell: every digital subsector is shrinking while in-person entertainment keeps growing, and Hollywood alone has lost over 100,000 jobs, a third of the sector, to its lowest employment since 2008. His worst case is the newspaper precedent, an 85% payroll collapse after the internet de-bundled the business model.

The Apollo note  ·  The apricitas essay  ·  HN: Apollo  ·  HN: apricitas

🪿 Goose: memory safety with no heap, no GC, no lifetimes

The HN thread spent the night arguing about a language whose entire memory model fits in one sentence: a handful of data stacks, and scope exit is the free. Goose is the latest language from Wouter van Oortmerssen (Aardappel), and its pitch is that memory safety does not need a garbage collector, a borrow checker or lifetime annotations. The compiler proves where everything lives; the runtime tracks only bump pointers.

The design is a set of interlocking constraints. At most one resizable value is live per stack and it is always on top, so growth never moves anything and never checks capacity, which is why a reference into a growing array stays valid forever, the thing C++ vector and Rust Vec cannot promise. Nested data stays inline: the demo order record with a customer string and items is 29 bytes and zero allocations, against 160 bytes and one allocation in C++ and 153 bytes and four in Rust. Threads share nothing: each worker compiles as a separate program and flat values cross typed queues as a memcpy. It compiles to a single C file, and the bundled TinyCC backend runs programs in-process with no build step. The author’s sixteen-benchmark suite claims 3.3x idiomatic C++, 1.16x hand-optimized C++ and 1.12x the best safe Rust, on 1.9x, 1.3x and 1.2x less memory. Those are his numbers on his suite; the losses are published alongside the wins.

The language site  ·  HN discussion

🐛 Tailworm: a three-day audit of Tailscale, published raw

The report opens with a researcher’s note that reads like a dare: “we have zero-trust in our infrastructure.” Erica Windisch of Cyberdione Labs spent three days this week reproducing attacks on the current Tailscale client and the self-hosted headscale control server in a lab, then published the full assessment with a legal argument in place of a disclosure timeline: the software is BSD-3, offered without warranty, so the audit treats users as fellow developers. Tailscale Inc. had no involvement and, as of this morning, no public response.

Five finding classes. The headline: Tailscale SSH’s unix-socket forwarding performs path checks and file operations as root, so races and a missing privilege drop turn one approved device plus one SSH policy rule into root execution on target servers, enrollment-key capture and arbitrary root-file deletion. Second: the experimental conn25 app connectors translate traffic to attacker-chosen destinations without consulting network ACLs, which the report maps to cloud metadata theft and Kubernetes service reach. Third: an internet-reachable headscale can be crash-looped without an account. Fourth: a malicious DERP relay can exhaust any client’s memory. Fifth: a serve TCP-forward to a unix socket can expose the node’s admin API. Their summary line: the mesh “converts one ordinary foothold into fleet-wide privileged access at machine speed.”

The report  ·  HN discussion

🧠 ThinkingCap teaches Gemma-4-26B to think 39% less

BottleCap AI keeps publishing the same trick on new bases: take a reasoning model, teach it to stop thinking sooner, change nothing else. The fifth ThinkingCap is their first on a Google model, gemma-4-26B-A4B-it, the sparse mixture-of-experts with 26B parameters and about 4B active per token, and they call it their cleanest result yet.

The headline numbers, all from BottleCap’s own harness: 38.9% fewer thinking tokens on average across twelve benchmarks, up to 69.1% on LongBench v2, for an average accuracy cost of 0.15 percentage points, smaller than the run-to-run spread on most of the suite. Looping responses drop from 3.64% to 0.76%, truncations from 4.35% to 0.72%. The methodology section is the part worth copying: multiple seeds per benchmark, 95% intervals, and the base model measured by them rather than quoted from someone’s leaderboard. The honest blemish: AA-LCR, a long-context retrieval test, gives back 4.0 points. Same sampling settings, same vLLM stack, weights and GGUFs on Hugging Face. A PRO tier claims up to 87.6% fewer thinking tokens, their number, behind an email address.

The release post  ·  HN discussion

🎬 video-editor-skills: Claude Code learns the rough cut

The repo’s pitch fits in a line: dump your footage, describe the video you want, and let Claude Code produce the rough cut in DaVinci Resolve. Not a new app, a skill pack: structured instructions that hand the agent an editor’s workflow instead of hoping it improvises one.

appariciojunior/video-editor-skills teaches transcript-driven cuts, vlog pacing and B-roll placement, riding the same agent-skills wave as the security-audit skill from No. 019. Three days old, 246 stars. Two caveats before you build on it: the license field is not set, and a skill is only as good as the model’s taste on your footage. But the category is the story: “skill” is quietly becoming the unit of packaged expertise for agents.

The repository

eMudhra: the deadline passed, and the regulator’s own site is revoked. No. 020’s CA compromise keeps compounding. Mozilla’s root-program bug sits at NeedInfo on eMudhra after it let the investigation deadline pass; India’s MeitY called the coverage social-media fear mongering; and captnemo’s scan now counts 1,377 Indian government domains on revoked eMudhra certificates, 811 still serving them, including the Controller of Certifying Authorities, eMudhra’s own regulator. Firefox refuses the sites via CRLite; Chrome loads them fine. Nemo’s scan →

Your phone, your rules: a lockdown countdown against a banking thaw. keepandroidopen.org is counting down about 70 days to the first enforcement wave of Google’s developer-verification program: from 2027, apps on certified devices need central registration with fee, ID and signing-key proof, which F-Droid calls existential. Counter-signal: about a dozen banking apps, historically hostile to GrapheneOS, just began explicitly permitting it. 530 and 628 HN points. Keep Android Open →

Roku moves its developer repos from GitHub to GitLab. The streaming company’s developer blog explains the move; 272 HN points of platform-loyalty discourse followed. Roku developer blog →

The PHP Foundation runs on one full-time staff member. Reuters reports the steward of the language that powers much of the web is “under-funded”, per its own director. Context for the week its flagship bug became a game show. Reuters →