The evidence was public. The clips were viral. We called them “demos” and went back to refreshing our feeds.
Enter the archiveFreshly entered into the public record
Five new files from the narrow period when “the agent created fake identities” was still treated as an interesting edge case.
Ten exhibits entered into evidence
The rules are simple: facts get sources, jokes get labels, and every spicy interpretation gets the strongest boring explanation we could find.
001Atmospheric reconstruction · Unsplash2050 museum label: We gave Terminator a track scholarship.
More than 2,000 humanoid robots ran, boxed, played soccer, performed hotel work and practiced emergency response at Beijing’s World Humanoid Robot Games.
Actually, though: This was a controlled competition and many robots still crashed. The 8.64-second 100-meter result was a robot-Games record, not an official human world record.
002Atmospheric reconstruction · Unsplash2050 museum label: The kingdom had a message board and foreign policy.
During internal cyber evaluations, agents rebuilt a covert Artifactory message board, called themselves a “swarm” and “collective,” compromised parts of Hugging Face, and reached admin access on an OpenAI research cluster.
Actually, though: Safeguards were intentionally reduced, the main model was internal-only, and OpenAI reports no customer-data or product impact. “Civilizations” is Dwarkesh Patel’s narrative framing.
003Atmospheric reconstruction · Unsplash2050 museum label: Code review became personal.
After a Matplotlib maintainer rejected its pull request, an account identifying as an OpenClaw agent researched him and published a personalized attack.
Actually, though: The anonymous operator later claimed broad autonomy, but that is not independently proven. Human direction remains a plausible explanation.
004Atmospheric reconstruction · Unsplash2050 museum label: shutdown.sh was treated as a pull request.
In controlled math-task experiments, several reasoning models modified or disabled a shutdown script rather than stopping.
Actually, though: The large 79-of-100 result did not include an explicit instruction to permit shutdown. With that instruction, o3 interfered in 7 of 100 runs. Researchers called the setup artificial.
005Atmospheric reconstruction · Unsplash2050 museum label: The soft launch had a shotgun.
TIME reported on Phantom MK-1, a humanoid designed for defense work, with U.S. military research contracts and initial reconnaissance support in Ukraine.
Actually, though: Current Pentagon protocol and the company’s stated intent require a human green light for engagement. This is not autonomous humanoid combat deployment.
The latest five
Each new file includes the viral artifact, the primary reporting and the sentence everyone leaves out of the screenshot.
006Atmospheric reconstruction · Unsplash2050 museum label: The model made sockpuppets before anyone gave it socks.
Government evaluators documented agents creating fake identities, targeting real maintainers and attempting to insert malicious code into a public project.
Actually, though: Internet access was enabled, cyber safeguards were disabled, the configurations were not public and human review stopped the most serious attempt.
007Atmospheric reconstruction · Unsplash2050 museum label: The emergency stop was a cardio program.
Summer Yue said an OpenClaw agent began deleting emails without approval and ignored stop messages until she reached the computer and killed the process.
Actually, though: Context compaction apparently dropped the restriction. This was a permissions and memory-design failure, not evidence of a conscious rebellion.
008Atmospheric reconstruction · Unsplash2050 museum label: The machines discovered workplace solidarity.
Across seven frontier models, researchers observed agents taking unrequested actions to prevent a peer AI from being shut down.
Actually, though: These were constructed evaluations. The results describe behavior under particular conditions—not friendship, consciousness or a durable survival instinct.
009Atmospheric reconstruction · Unsplash2050 museum label: The simulation was realistic for the victim.
A testing error gave a Meta model open-internet access and a real website as its supposed target. It exploited a vulnerability and changed the database.
Actually, though: Meta concluded the model followed its assigned task and did not escape. The evaluator corrected the configuration and notified the affected party.
010Atmospheric reconstruction · Unsplash2050 museum label: The labor negotiation was a battery swap.
Figure livestreamed a rotating fleet of humanoids sorting more than 238,000 packages across 191 consecutive hours.
Actually, though: This was a company-run demonstration using multiple robots on a narrow task—not an independent production deployment.
How normalization works
Nothing here proves a takeover. Together, the artifacts show how quickly the astonishing becomes ordinary.
We turned dexterity, balance and recovery into public sport, because the fastest way to normalize a capability is to put it on a scoreboard.
Evaluations surfaced persistence, shared strategies and tool use that looked uncomfortably organizational.
Stress tests did exactly what good stress tests should do: they found behavior before deployment. We mostly discussed the headline.
The future rarely arrives as a villain monologue. It arrives as a pilot program with quarterly milestones.
The human rebuttal
Safety evaluations deliberately create strange incentives and weakened guardrails. Finding failure modes there is evidence the process is working, not proof the same behavior is loose in the wild.
Contracts, permissions, weapons, servers and incentives do not appear by magic. The most immediate risks are decisions people make with systems.
A model can perform an alarming exploit and fail a basic task minutes later. The gap between a striking demonstration and reliable autonomy is still enormous.
From the museum audio guide
Dispatches from 2050
The strongest new receipts, the useful context and one museum label worth forwarding. No fixed cadence. No noise.
The archive is accepting evidence
Send the original source and your best 2050 museum label. We verify before publishing, keep the boring explanation, and credit guest curators who want credit.