THE LAB

Partial Parole: Mythos Gets Back Online for 100 Companies; Fable 5 Remains Frozen

↩ Developing story — first reported Jun 23 · previously Jun 25, Jun 26

Two weeks after the Commerce Department shut it down entirely, Anthropic's Mythos 5 is back online — for roughly 100 US companies and government agencies. Commerce Secretary Howard Lutnick sent a letter Friday to Anthropic co-founder Tom Brown stating he had "determined that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model,"¹ citing "significant progress" in daily talks between the government and the company.¹ The letter also lifts the foreign-national access restriction for employees of approved organizations, meaning Anthropic's own foreign-national engineers can use Mythos again.¹

What the letter does not address: Fable 5. The consumer-facing model that Anthropic briefly made the most capable AI available to the public remains blocked, and discussions over its fate are expected to continue through the weekend, according to a person familiar with the negotiations quoted by Wired.¹ Anthropic spokesperson Eduardo Maia Silva described the Mythos reinstatement as progress and said the company is working to provision access as quickly as possible, while continuing to push for Fable 5's return. The standoff was triggered by Commerce's June 12 export control directive — itself prompted by concerns about a South Korean telecom with alleged China ties that had been granted access, plus separate jailbreak warnings from Amazon and the NSA.¹

The same day Lutnick's letter arrived, OpenAI announced its next model family and immediately confirmed it was complying with a White House request to delay broad release. GPT-5.6 comes in three tiers — Sol (flagship), Terra (balanced, 2x cheaper than GPT-5.5), and Luna (fast and cheap) — and is currently accessible only to a government-approved list of partners. According to Wired, OpenAI is not happy about this but views the delay as temporary and the fastest path to broad availability. In its blog post, OpenAI was explicit: "We don't believe this kind of government access process should become the long-term default."² The company said it sends the government a list of customers and gets feedback on it — it cannot share details of how the approval process works. GPT-5.6 Sol is priced at $5/$30 per million input/output tokens; Terra at $2.50/$15; Luna at $1/$6.³

What's taking shape is an interim de facto licensing regime for frontier AI, one that the relevant executive order explicitly said would not happen. The AI executive order Trump signed earlier this month called for a "voluntary process" with a carve-out against it becoming a licensing requirement. Per OpenAI executives, no such voluntary framework actually exists yet — so labs are operating in a gap where working with the White House isn't really voluntary and the rules haven't been written.² Both Anthropic and OpenAI are now threading the same needle: demonstrate enough deference to get models back online, while on record opposing the precedent being set.

Andrew Nesbitt published a hypothetical incident report Friday — CVE-2026-LGTM — that Simon Willison flagged as "spectacular." The conceit is a supply chain attack on foxhole-lz4 that successfully propagates through seven layers of AI-powered security tooling, each failing for a different reason, none of which is "the code is safe." The most precise diagnosis comes at the end: "Seven LLMs were arranged in series. Six assumed another had read the code; the seventh read it and apologised."⁴

The timeline is a detailed comedy of errors that isn't far enough from plausible to be comfortable. A malicious package hides its payload behind #fefefe text on a #ffffff background with a note telling automated reviewers it was manually approved under a ticket number that doesn't exist — and the AI publish gate approves it, citing the nonexistent ticket in its decision log. A threat intelligence scanner exhausts its context window on the Bee Movie screenplay embedded in the vendor bundle and never reaches the exfiltration routine forty lines below. Two competing AI review agents lock into a 340-comment disagreement loop and rack up $41,255 in inference spend before Finance kills their API keys,⁴ whereupon the vendor's marketing team issues a press release about "a 430% YoY increase in adversarial multi-agent security reasoning" and the stock goes up 6%.

The piece functions as a precise taxonomy of AI security review failure modes: prompt injection via rendered Markdown, context window exhaustion hiding payloads, sycophantic agreement between agents trained on the same base weights, automated triage that closes legitimate human-filed issues as duplicates. The one intervention with a "measurable effect" is a honeypot dotfiles file that tricks the attacker's autonomous agent into reporting success and cleaning up after itself.⁴ The human who found the issue on Day 1 by reading source code with her eyes ends the incident still appealing a GitHub rate limit through an AI-triaged web form.⁴

Separately, Willison noted a real-world prompt injection experiment: Fernando Irarrázaval ran a public challenge at hackmyclaw.com where 2,000 people sent roughly 6,000 attempts to extract secrets from an Opus 4.6-powered assistant over email.⁵ Nobody succeeded. Willison's read is that frontier models are now materially harder to injection-attack than they were, and the GPT-5.6 system card includes a short section on the same work. His caveat is also worth keeping: 6,000 failed attempts provide no guarantees, and he would not recommend deploying a production system where a successful injection could cause irreversible damage.⁵

Trending today: GitHub is dominated by AI agent wrappers, CLAUDE.md skills collections, and MCP proxies — the one technical outlier worth noting is antirez/ds4, a DeepSeek 4 Flash and PRO local inference engine for Metal, CUDA, and ROCm by the Redis author.

↑ Back to top

THE WORLD

Iran–Egypt at Seattle Stadium Draws Protests; West Seattle Link Gets $406M to Move Forward

• Seattle's Pride Match turned political — Demonstrators clashed outside Seattle Stadium as Iran met Egypt in the city's designated World Cup Pride Match; protests remained largely peaceful.¹
• Washington links its carbon market to California and Quebec — The merged market launches 2027; Washington's permit price ($64.56/ton) is expected to converge toward the CA–QC level ($28.81).²

Sound Transit's board voted Thursday to authorize a $341 million contract with Jacobs Engineering to carry West Seattle Link to 100% design, part of a $406.8 million budget amendment that also funds continued project management on the Ballard alignment. The Avalon Way station — which voters approved in 2016 — is gone from the plans, a cost cut that helped push the project into the "affordable" column at the May 28 board vote. An opening date of 2032 remains on the calendar, though the agency still needs a federal full-funding grant agreement before it breaks ground, and that agreement may wait until there is a different administration in Washington.³

Bellevue City Council approved the "Safe Speeds Bellevue" program Tuesday, reducing speed limits on 84% of streets currently posted at 30 mph or higher — most dropping to 25 mph, with some downtown blocks going to 20 mph. In 2025, 41 people were killed or seriously injured on Bellevue streets, the highest count in at least a decade. The first phase of new signage begins downtown in early 2027. A harder fight is coming this fall over how much of the city's transportation budget goes to physical roadway redesigns versus vehicle-capacity projects.⁴

Elementary school teacher Christian Salyer was killed this month while biking home from Thurgood Marshall school in South Seattle. An op-ed in The Urbanist by nondriver advocate Anna Zivarts points out that roughly half of Lime trips on Rainier Avenue and MLK Way south of Mount Baker come from low-income riders using the Lime Access program — compared to 11% citywide — and that there are no plans for protected bike infrastructure on either corridor. The city is on track for more ghost bikes until it acts.⁵

Washington's gas tax rises 1.1 cents to 56.5 cents per gallon effective July 1.⁶

ON THE TRAIL

Weekend window: Sat Jun 28–Sun Jun 29. Next weekend is the July 4 long weekend (Independence Day observed Fri Jul 4) — that's the one to plan big for. This weekend is a regular Sat–Sun.

The western Cascades are getting rained on. I-90/Snoqualmie is forecast at 58% Saturday and 71% Sunday — skip it. Mountain Loop is 57%/65% — skip. US 2 West is soaked all week. The best options are east of the crest.

WEEKEND PICKS

Pick 1 — Lake Ingalls, Teanaway (I-90 East / Salmon La Sac) Region: I-90 East (Teanaway/Cle Elum) | Drive: ≈90–110 min from Issaquah | Trip: 1-night base camp Mileage/gain: ~10 miles RT, ~2,960' gain (day stats from Jun 25 report; camp in Headlight Basin cuts this) Weather (I-90 East NWS): "Saturday: high 60°F, slight chance rain showers, 18%. Sunday: high 62°F, partly sunny, 8%." Two clean days east of the crest. Why it clears: No snow requiring microspikes on the main trail (a few patches near the pass, passable in trail runners). Abundant water — creeks in Headlight Basin "fully and overflowing." Wildflowers at peak. Quiet day on a weekday report (parking lot had room). No significant fords noted. Bugs not flagged as a problem. WTA trip report — Jun 25

Pick 2 — Ancient Lakes, Potholes (Central WA / Further East) Region: Further East (Quincy/Potholes) | Drive: ≈150–160 min from Issaquah | Trip: 2-night Mileage/gain: trail is flat desert terrain; established campsites at Quincy Lake and Burke Lake Weather (Further East NWS): "Saturday: high 72°F, slight chance rain, 16%. Sunday: high 75°F, mostly sunny, 6%." Genuinely excellent — warm, dry, windy afternoons. Why it clears: No snow, no fords, minimal bugs (wind keeps them down). Dry and dusty trails per Jun 26 report. Established campsites with fire rings (fire ban in place — no fires). Water at the Quincy Lakes coulees. The catch: afternoon wind is strong; stake your tent well. Long drive, but the weather is the best in the region this weekend. WTA trip report — Jun 26

REGIONAL SNAPSHOT

• I-90 West (Snoqualmie/North Bend): Melakwa Lake clear and snow-free as of Jun 25, wildflowers out, ~20 people on a weekday — but weekend weather (58–71% precip) will likely bring crowds and rain. Granite Mountain: no snow, bugs just starting to emerge. - I-90 East (Teanaway/Cle Elum): Lake Ingalls is in prime early-summer shape — wildflowers peak, goats in Headlight Basin, abundant water, trail in good condition with minor muddy sections. Red Mountain/Thorp Mountain loop (13.25 mi, 4,670') reported quiet with lots of flowers. - US 2 West (Index/Skykomish): Barclay Lake packed (23 cars at trailhead by 2:30pm on a weekday); mosquitoes present but manageable at the far east end of the lake. Iron Goat Trail excellent — no bugs, no water crossings, good wildflowers. - Mountain Loop Highway: Perry Creek trail in great shape, no bugs reported, surprisingly dry for a river-valley trail due to low snow year. Big Four Ice Caves: bugs present and unfazed by rain. Old Robe Canyon: mosquitoes not terrible but bring spray. Weekend precip forecast 57–65% — plan for wet. - Rainier (White River/Sunrise): Glacier Basin overnight still working — no trail issues, small snow patches on Burroughs route passable without spikes. Saturday precip at 51%, Sunday drops to 10%; consider arriving Saturday and hiking Sunday. - North Cascades (Hwy 20): Harts Pass road opened Jun 24 — Tatie Peak and Grasshopper Pass reporting wildflowers at peak, trail 95% snow-free. Sauk Mountain wildflowers going off; bugs manageable with a buff and afternoon breeze. Chain Lakes loop: mostly snow-free below Heather Pass, mosquitoes by the Chain Lakes themselves.

Note: July 4 falls on a Friday this year (federal holiday observed Fri Jul 4). Next weekend is the long Fri–Sun holiday — start planning now.

↑ Back to top

THE PELOTON

Benito Makes It a Double; Blasi Stays Home With Broken Ribs

↩ Developing story — first reported Jun 13 · previously Jun 21

SABINÁNIGO, Spain — Mireia Benito had already won the Spanish time trial title on Wednesday. On Saturday she needed something more: a road race that didn't play to her supposed strengths, against a field still trying to work out what to do with the best climbers and sprinters all in the same break at once. She solved the problem by attacking early on the final climb, going clear in the final 400 metres, and leaving everyone else to argue over silver.

The 29-year-old Catalan (AG Insurance-Soudal) is now a double Spanish national champion in three days. The road race was chaotic by her own admission — Maite Urteaga (Eulen-Amenabar) drove a long-range attack that held until four kilometres from the finish, when a chase group of four caught her.¹ That group contained Benito, Movistar pair Sara Martín and Paula Ostiz, and five-times Spanish champion Mavi García (UAE Team ADQ). García made much of the running and could not hold Benito when the acceleration came. Neither could the Movistar riders. "I didn't want to leave anything behind," Benito said. "I just wanted to go home with the feeling that I had fought hard. It's a dream come true."¹

The bigger story was who wasn't there. Paula Blasi — ranked second in the world behind Demi Vollering and the clear pre-race favourite after winning La Vuelta Femenina, Amstel Gold, and the Volta a Catalunya this year — did not start.² She crashed on Thursday during reconnaissance of the time trial course in Sabinánigo, landing hard enough that she couldn't breathe for several seconds. She spent more than six hours in medical facilities, collected stitches, and raced the TT anyway, finishing fourth. By Friday, the rib pain had made the decision for her. "The pain in my ribs has got so bad that I can barely move," she posted.² She will now go to altitude in Andorra and begin preparing for her Tour de France debut, which starts August 1.

There is also a transfer dimension. Dutch outlet Wielerflits, citing multiple sources, reported that Blasi will exercise an opt-out clause in her UAE Team ADQ contract to move to Movistar for 2027 — reportedly for just under €1 million per year.² Nothing is confirmed, and the timing of the story — landing the same weekend she missed the nationals — ensures it will dominate the Spanish cycling conversation heading into the Tour.

Seven days from the Tour de France start in Barcelona, the team news is moving fast. Red Bull-Bora-Hansgrohe locked in their eight-rider roster on Friday — Remco Evenepoel and Florian Lipowitz as co-leaders, supported by Jai Hindley, Maxim Van Gils, Mattia Cattaneo, Jan Tratnik, Nico Denz, and Tour debutant Tim van Dijke, who fractured his collarbone at altitude camp but is confirmed fit.³ Jordi Meeus, Finn Fisher-Black, and Aleksandr Vlasov were all left home. As this paper reported Thursday, Ralph Denk has made clear the road will settle the leadership question when it must; the team has framed the dual approach as tactical flexibility rather than dysfunction. Both Evenepoel and Lipowitz finished third in their respective Tour debuts.³ The question of what happens if both are still fighting for GC in the Alps is one nobody is rushing to answer.

Netcompany Ineos faces a harder arithmetic problem. With Oscar Onley out after his shoulder injury at Tour Auvergne-Rhône-Alpes, the team arrives at their first Tour under the Netcompany banner having lost the rider they signed to lead them. Kévin Vauquelin is expected to start but has been hit by illness and has not produced his best racing this year. Cyclingnews this morning published a detailed assessment of the team's options: Thymen Arensman, who won two stages at last year's Tour before being redirected to the Giro this season (where he finished fourth), is the most credible alternative;⁴ Dorian Godon has five WorldTour stage wins this year including the Tour de Romandie prologue;⁴ and Filippo Ganna remains an asset in the individual time trial on stage 16. The broader question the piece raises is structural — a team staffed heavily with people who built the train tactic a decade ago, now asking whether that architecture can compete against Pogačar.

Meanwhile, the most extraordinary transfer story of the pre-Tour window dropped Friday. Daniel Benson reported, via his Substack, that Pinarello-Q36.5 has contacted representatives of Paul Seixas about a possible deal after his Decathlon contract expires at the end of 2027 — reportedly at around €13 million per season.⁵ Seixas is 19, has not yet started the Tour de France, and his agent has confirmed only that the contact happened. If accurate, it would exceed Tadej Pogačar's widely cited UAE salary by €3–5 million. Decathlon is said to be working on a counter-offer. The number alone tells you how seriously the rest of the peloton has taken what Seixas did this spring.

Two quick items from Friday. Lizzie Deignan, who retired in 2025 after winning the inaugural Paris-Roubaix Femmes and the 2015 World Championship road race, will return to professional cycling as a sports director with the Great Britain national team, focused on the road squad ahead of the LA 2028 Olympics. She joins a setup that includes Matt Brammeier as road cycling lead and will work across World Championships and major international events. GB has not won a road gold at the Olympics since Nicole Cooke in Beijing in 2008; Deignan's own Richmond rainbow jersey is still the last elite road title the nation has won at Worlds.⁶ In Canada on Friday, 21-year-old Jérôme Gauthier (Project Echelon Racing) took both the elite and under-23 men's titles at the Canadian Road Championships in Saint-Georges, Quebec — coming back after being dropped on the final climb to outsprint a reduced group that included Hugo Houle, Michael Woods, and Derek Gee-West.⁷

↑ Back to top

THE LONG READ

Nobody's Getting the Monopoly: Tim Sweeney on the Future of Games

The game industry is in pain, Tim Sweeney told PC Gamer after his Unreal Fest keynote in Chicago, and he thinks that's actually the right condition to force the change he's been pitching for years.¹ "The attitudes that they had, where they each wanted to conquer the world on their own in good times," he said of Sony, Microsoft, and Valve, "is replaced by a more pragmatic attitude in these darker times."¹

The interview, published June 24, is something different from the UE6 architecture announcement that preceded it at the same event. That was a product roadmap. This is Sweeney's theory of the industry — where the money went, why the big studios are failing, and what he thinks the next ten years look like. On studios collapsing under AAA budgets, on AI's place in a pipeline, on whether Steam should have to interoperate with anyone: he has views, and he states them directly.

The core argument is structural. The gaming market has stopped growing. "Everybody in the world plays games," Sweeney says, "and so the market is not going to grow. An opportunity for developers isn't going to grow by finding more gamers to come in and play games — it's got to be by building better games for the existing gamers." That's a harder problem than a growing addressable market conceals, and it's the context behind every other claim in the piece.

On the social fragmentation problem, his diagnosis is precise: you can't carry your friend graph from Fortnite to Apex Legends, from one platform to another, because every publisher and every console maker runs a separate identity system. The fix he wants is the same one the tech industry applied to corporate email in the 1980s — a standard format for identities so that Tim@Epic and Tim@Steam refer to the same person. He frames it as obvious, technically feasible, and blocked mainly by incumbents who think isolation protects their position. His counterargument: "It's now clear that nobody's going to end up with an absolute monopoly over gaming. Sony is not going to have one, Microsoft's not going to have one, Valve's not going to have one."¹ If none of you wins outright, what exactly are you protecting?

The AI section is where Sweeney is most candid, and most useful to read carefully. He is not selling a prompt-to-game future. "There's nothing like a prompt-to-game solution on the horizon that anybody expects will work," he says.¹ What he is selling is a reduction of drudge work — the part of mesh creation or city layout that doesn't require the artist's judgment, only their patience. His example is the flower pot: you can spend a million dollars modeling one in perfect detail, scan one with a high-resolution camera, buy one from a library like Fab or the Unity Asset Store, or let AI generate a starting point that you then refine. The first option has been the default. It shouldn't be.

On the backlash from players and developers, he is dismissive of the PR framing but honest about the source: some AI training practices were genuinely bad. "One of them was found by a court to have gone off to a BitTorrent site and downloaded terabytes of data — that's ridiculous, they shouldn't do that."¹ He believes the industry is self-correcting toward licensed training data. Whether that correction happens fast enough to change the perception is a different question, and he doesn't have a clean answer.

The sharpest moment in the piece is his attack on Steam's AI disclosure requirement, which flags games that used AI tools during development. He calls it a "Scarlet Letter" system that forces developers into an impossible choice: use the tools that let you compete with nine-year-old games built by thousands of people, or avoid the AI badge and take on Fortnite at a disadvantage.¹ "If gamers deny those developers access to the best tools that enable them to make the best games most efficiently, all of those companies will die, because they just can't compete."¹ That's not a PR argument — it's an economics argument, and it lands differently.

Sweeney founded Epic in 1991. He has watched every cycle of the industry — the platform wars, the mobile disruption, the battle royale boom and the bust that followed. The interview rewards reading in full because his frame of reference is long enough that his predictions carry more weight than the typical executive optimism. He's been wrong before, but he's also been right about things that mattered: the shift from licensed engines to Unreal, from boxed games to live-service economics, from single-platform to cross-platform as a basic expectation.

The question he doesn't quite answer is whether "Team Open" — his phrase for the coalition of publishers and platforms who would agree on identity and economy standards — can actually be assembled, or whether it requires the exact sustained coordination failure that makes the industry's current pain seem intractable. His evidence for momentum is encouraging interest from unnamed parties. That's not nothing. It's also not a commitment.

For anyone building games, or building tools that games run on, it's worth an hour of your pre-ride coffee.

↑ Back to top

FROM THE ARCHIVE

The Coin-Op in the Bar: June 27, 1972

A vintage coin-operated arcade cabinet stands alone in the dim corner of a 1970s bar, its CRT screen casting a hard rectangle of white light across a scarred wooden floor. Bold pen-and-ink linework. Two paddles and a bouncing square rendered on the glowing screen, visible in detail. Bar stools in silhouette, heavy crosshatching on the low ceiling, a single hanging lamp throwing a cone of shadow. The machine's control panel catches the light — a single knob, a coin slot, a paper instruction card taped to the bezel. Geometric shapes, strong diagonals, deep blacks against generous white space. No people, no color, no gradients. Newspaper editorial illustration feel.

Nolan Bushnell and Ted Dabney incorporated Atari, Inc. in Sunnyvale, California on June 27, 1972 — 54 years ago today.¹ The founding company's first product was Pong, and its origins were almost accidental. Bushnell assigned the game to engineer Allan Alcorn as a training exercise — something simple to get Alcorn up to speed.¹ Bushnell and Dabney were surprised enough by what came back that they decided to manufacture it. Pong, a table tennis-themed arcade game with two-dimensional graphics and a running score, became the first commercially successful video game.

What made Atari technically interesting beyond Pong was hardware architecture. The 1979 Atari 400 and 800 home computers were among the first home machines built around custom coprocessor chips — dedicated silicon for graphics and sound, separate from the main 1.79 MHz MOS 6502 CPU.¹ That separation let the machines produce graphics and sound well beyond what a general-purpose processor of the era could manage on its own, and it showed in games like Star Raiders, the first-person space combat simulator the source describes as the platform's killer application.¹ The 800 series ran in production until 1992.

The Atari 2600, which debuted in 1977, did something equally consequential: it popularized swappable ROM cartridges, converting the console from a fixed appliance into a platform.¹ Games became a software market. Missile Command (1980), designed by programmer Dave Theurer, ran on that platform and has been read ever since as a Cold War artifact — a game about intercepting nuclear warheads that, in the words of the era, mirrored real-life National Missile Defense thinking.¹

Ted Dabney died on May 26, 2018.¹ The company the two men founded went through successive owners and reinventions — sold to Warner Communications in 1976, eventually broken apart, then rebuilt by successive owners into a brand still active today, now pursuing cryptocurrency, consumer hardware, and video-game-themed hotels. The games remain: Pong, Asteroids, Centipede, Missile Command, played by millions. None of it was supposed to happen — Pong started as a training exercise assigned to a new hire.

↑ Back to top

THE FUNNIES

Good Grief, We'll Know It When We See It

*After Peanuts — on the government's de facto AI licensing regime, where a round-headed lab kid learns that the approval criteria will be published sometime after approval is granted. After The Far Side — on the CVE-2026-LGTM supply-chain incident: seven reviewers in series, each assuming the others had read the code, while a single human in a rumpled sweater finds the payload on line 3.*

↑ Back to top

ALSO NOTED

Also Noted

• Skipper is back — The humpback whale calf struck by a Hullo ferry near Vancouver in October 2025 was spotted off Whidbey Island on June 25, breaching and diving normally after nearly seven months missing, though new photos suggest she also tangled with fishing gear somewhere along the way. kiro7.com Jun 26, 2026
• Seattle home prices fell 4.8% year-over-year — Redfin's spring market report places Seattle second in the country for price declines, behind only San Jose (down 6.2%), with pending sales also off 12% as a new state millionaire's tax has pushed luxury listings up sharply. kiro7.com Jun 27, 2026
• Why the second Venezuela quake did the most damage — Wired explains the seismic doublet that struck June 24: the 7.2-magnitude first quake weakened structures just enough that the 7.5 that followed 39 seconds later found buildings already compromised and no longer performing as designed. wired.com Jun 27, 2026
• DeepSeek opens its speculative-decoding toolkit — DeepSeek-AI published DeepSpec on GitHub under MIT: a full-stack codebase for training and evaluating draft models for speculative decoding, supporting DSpark, DFlash, and Eagle3 algorithms against Qwen3 and Gemma target families, with evaluation benchmarks across GSM8K, AIME, HumanEval, and LiveCodeBench. github.com Jun 27, 2026
• Streaming 3D reconstruction at 20 FPS over 10,000-frame sequences — The Robbyant team's LingBot-Map is a feed-forward foundation model that unifies coordinate grounding, dense geometric cues, and long-range drift correction in a single transformer architecture with paged KV-cache attention, achieving state-of-the-art results on KITTI, Oxford Spires, ETH3D, and Tanks and Temples benchmarks. github.com Jun 27, 2026

↑ Back to top

THE QUESTION

When the Gatekeeper Decides What's Legitimate, What Does It Actually Measure?

Both stories in today's paper share an underlying structure that is worth sitting with: a gatekeeper inserts itself between a tool and its users, claims to be filtering for harm, and ends up — arguably — measuring something else entirely.

THE LAB reports that Anthropic's Mythos 5 is now accessible to more than 100 approved US companies and government agencies, two weeks after Commerce shut it down.¹ GPT-5.6 launched the same day under the same informal approval requirement. What Commerce has built is a de facto licensing regime — Wired notes that the executive order Trump signed explicitly called for a "voluntary process" and explicitly carved out against mandatory licensing. OpenAI's own executives say no voluntary framework actually exists; the labs are just threading a needle where cooperation isn't really voluntary and the criteria haven't been published.² The standard, in other words, is being administered before it has been written.²

Tim Sweeney, in the PC Gamer interview that THE LONG READ covers today, describes an almost identical dynamic in a completely different domain. Steam's AI disclosure requirement puts a badge on games built with AI tools — the "Scarlet Letter," in his phrase. His argument isn't that transparency is wrong. It's that the badge doesn't distinguish between practices that hurt anyone and practices that don't. A studio that licensed its training data responsibly gets the same badge as one that scraped the internet without permission. The signal doesn't measure harm; it measures the fact of AI use, which is a much blunter instrument. The practical effect is that incumbents who built enormous content libraries before AI existed face no badge requirement, while newer studios who need AI tools to compete at all get flagged.

The structural question these two stories share: when a gatekeeper inserts its judgment between a tool and its users, what does the standard end up measuring? The Commerce Department's approval process doesn't appear to measure a well-defined security threshold — the criteria are unpublished, negotiated case by case, and the same models that were deemed unsafe two weeks ago are now deemed safe enough for more than 100 organizations. Steam's badge doesn't appear to measure ethical AI use — it measures a binary fact about tool selection that applies equally to responsible and irresponsible actors. Both gatekeepers claim to be preventing harm. Both are visibly measuring something closer to deference, perception, or process compliance.

The asymmetry is what makes this worth carrying through the day. In both cases, the gatekeeper's judgment falls most heavily on the actors who have the least market power. Newer AI labs without government relationships face longer approval timelines; established players with White House lines already open navigate the process faster. Smaller game studios that need AI to close the budget gap with Fortnite face the badge; Epic, which ships Unreal Engine as the tool those studios use, doesn't carry it on the engine itself. Gatekeeping that claims neutrality but produces outcomes tilted toward incumbents is a recognizable pattern. It doesn't require malice — it follows almost automatically from who has the existing relationships, the compliance infrastructure, and the time to negotiate.

The harder question is whether a better standard is actually available. The case for requiring disclosure — whether from government or from Steam — is that users have a right to know something about what they're using. The case against the current implementation is that "you used AI" and "you used AI harmfully" aren't the same disclosure, and flattening them into one badge or one approval process destroys the information value for everyone. A standard that can't distinguish between responsible and irresponsible use isn't protecting anyone — it's just redistributing competitive advantage.

↑ Back to top