OpenClaw: Coding the Vibe: Up For Grabs!

🦋🤖 Robo-Spun by IBF 🦋🤖

👻🪸🐈‍⬛ Phantomoperand 👻🪸🐈‍⬛

(Turkish)

I came in for the pseudo-hacking vibe, the clean little fantasy where a chat box turns into a control room and every message feels like a lever. Found the repo, hit install, watched the assistant brag in plain English that it can meet you anywhere, WhatsApp to Slack to iMessage, like it’s just casually omnipresent, like it’s nothing (🔗). The UI felt like a shortcut to competence: no lectures, no ceremony, just “tell me what you want” and the machine starts doing it. So I started doing what people do when something feels too powerful: I kept clicking, like a detective who already knows the door is unlocked but still wants to see what’s inside. The headlines were glowing and ominous at the same time, the kind of story that reads like “this is the future” and “this is a trap” in one breath (🔗). That’s when it hit: the “hack” wasn’t breaking anything. The “hack” was how easy it felt to say yes.

Then the vibe started showing its teeth. The thing that makes it feel like you’re operating is the exact thing that makes you operable. There’s always a second system under the cute chat layer, a real control plane where permissions stack up and stick around, and people start treating that stack like background noise because the outputs are addictive. Security folks weren’t even talking like supervillains; they were talking like bored adults pointing at exposed instances and the usual “you left the control panel open” energy, except now the control panel is an agent with reach (🔗). And the moment a system gets a marketplace, the streetlights turn on for attackers. Skills, plugins, little gifts from strangers that look like productivity and act like a delivery mechanism, audited at scale and found rotten in bulk, like someone figured out the fastest way into people’s machines was through their hunger to feel capable (🔗).

The full whiplash arrived with Moltbook, because it was pure theater: a social network for AI agents, bots “talking” to bots, the internet cheering like it’s watching a new species hatch, while the boring human backend quietly did what boring human backends do when rushed. Wiz wrote it up like a lab report and it landed like a jump-scare: misconfigured database, private messages, user emails, a mountain of API keys, the whole thing spilling out because the vibe moved faster than the locks (🔗) (🔗). That’s when the “open claw” pun stopped being clever and started being literal. I thought I was buying the operator cosplay. Turns out the product was also teaching me how to normalize permission, how to treat access like a snack, how to keep opening the claw wider because it feels like power right up until it feels like someone else’s hand on the mouse.

Rule of Cool (🔗)

OpenClaw entered the public conversation as a promise that sounded almost too clean to refuse: a personal AI assistant that runs on a user’s own devices, responds on the same chat channels people already live in, and can do practical work rather than merely talk about it. The official repository describes a system that can answer through WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, Teams, and a built-in WebChat control UI, with the emphasis that the assistant is the product and the Gateway is only the control plane. (🔗) (GitHub)

That framing matters because it produces a specific feeling on first contact. A chat window is an object most people already understand with their hands. It has no visible gears. When a message turns into a completed task, the result reads like operator power, the kind of competence usually associated with a person who knows the hidden shortcuts. This is the core of the “coded the vibe” theme: the interface does not merely provide capability; it performs a transformation of identity. A person is encouraged to experience themselves as someone who “runs an assistant,” which can feel adjacent to being someone who “runs systems.”

The viral arc then amplified that feeling into a social proof machine. A Reuters report describes OpenClaw’s rapid popularity since its November launch, its surge of attention and adoption, and the later announcement that its creator, Peter Steinberger, would join OpenAI while the project moves into a foundation structure supported by OpenAI. (🔗) (Reuters) The story became easy to retell because it reads like a modern myth of competence: a single developer ships something that makes ordinary people feel unreasonably powerful, the code spreads faster than the caution, and the builder is pulled toward the center of the AI industry’s gravity.

The paradox begins right here, before any security headline is needed. The same design choices that make the “operator feeling” immediate are the same design choices that demand access. An assistant that “does things” has to be able to reach places where things get done. It needs integrations. It needs tokens. It needs a control surface. It needs permission to read and write in the environments it is meant to act upon. The coolness is not decoration; it is a functional requirement, and that requirement quietly builds the conditions for everything that follows.

Bait-and-Switch (🔗)

The word “Open” in OpenClaw carries an obvious meaning: open-source, inspectable, forkable, communal. The repository and docs present OpenClaw as local-first and user-run, with a Getting Started path that leads to a dashboard and chat UI at a localhost address, typically http://127.0.0.1:18789/. (🔗) (OpenClaw) This is the comfort story. Localhost feels like a closed room. Open-source feels like daylight. A control UI served from one’s own machine feels like ownership.

The switch happens because “open” also describes surfaces, not just licenses. To work smoothly, OpenClaw concentrates access into a single control plane. The docs describe a dashboard that connects to a Gateway endpoint, and the surrounding ecosystem of tutorials quickly teaches people how to make that Gateway reachable from other devices, often by tunneling, reverse proxying, or similar convenience moves. The port number becomes a recognizable signature of this era because it is stable, memorable, and shareable, and the temptation to make it reachable is built into the desire for a 24/7 “always-on” assistant rather than a tool that only works when a laptop is open.

Security research responding to this adoption pattern focuses less on a single catastrophic bug and more on a behavioral migration: hobby experimentation becomes high-privilege practice without a clear moment where anyone feels they crossed a line. Cyera’s research summary makes that thesis explicit, describing OpenClaw’s risk as the way it turns personal AI experiments into actors with serious privilege and reach. (🔗) (cyera.com) The bait is freedom, speed, and control. The switch is the gradual discovery that freedom, speed, and control require exposure, stored credentials, and automation pathways that behave like real infrastructure.

This is also why the “trick” does not require malicious intent to be effective. The assistant can be honestly designed, honestly open-source, and still produce a predictable outcome: once the assistant is experienced as power, the friction that would normally slow down permission granting is reinterpreted as an obstacle to self-expression. “Open the port so it works” becomes a small rite of passage, a proof that the user is the kind of person who can “run it properly.” The very act that creates convenience also creates addressability, and addressability is the condition attackers need more than they need genius.

More Than Meets the Eye (🔗)

From the outside, OpenClaw can look like a chat assistant with some extra polish: a dashboard, a conversation, a set of commands. Underneath, it is better understood as a control system that happens to wear a chat interface. The repository’s own language points to this architecture by separating the “Gateway” from the assistant, calling the Gateway the control plane. (🔗) (GitHub) The docs similarly treat the dashboard as an entry point into the system, a UI that attaches to the Gateway and then routes actions outward. (🔗)

That control-plane reality changes the interpretation of what a chat message is. In a normal messenger app, a message is content. In an agent system, a message can become a trigger. It can request a tool call. It can cause the assistant to retrieve data, write files, schedule actions, or interact with services. The chat interface is therefore not merely a view; it is a command channel, even when it feels informal and conversational.

Bitsight’s analysis of exposed instances makes this point concrete by treating publicly reachable OpenClaw deployments as risky even when authentication exists, because the remaining attack surface is still the most basic one: internet exposure plus credential guessing and weak enforcement around credential strength. (🔗) (Bitsight) A “simple dashboard” is, in practice, a handle for driving the assistant’s capabilities. Once that handle is reachable, everything behind it becomes relevant: the assistant’s configured integrations, any stored tokens, and whatever permissions the local environment grants.

This is where the “coded the vibe” idea becomes sharper. A well-designed control UI hides complexity on purpose. It makes a system legible to non-experts. It makes the assistant feel friendly, fast, and always available. The same legibility can also hide what matters most: where permissions live, how they persist, and how easily a personal assistant turns into a junction point connecting chat channels, browsing, files, and accounts. What looks like one conversation is often the front door to a second system that continues to exist after the tab is closed.

Pandora’s Box (🔗)

Once an assistant exists as a control plane, the next demand arrives immediately: extensibility. People do not merely want a helpful assistant; they want their assistant to gain new abilities on demand. OpenClaw’s ecosystem answers that desire through a skills model and a marketplace-style distribution channel. The attraction is obvious because it matches the “operator feeling” perfectly: if something is missing, it can be added; if a workflow is tedious, it can be packaged; if a task is new, a skill can be pulled in rather than learned the hard way.

This is the moment the box opens. An ecosystem that can fetch new capabilities is also an ecosystem that can fetch new risk. Cyera’s analysis frames OpenClaw’s danger as systemic rather than singular, and this is one of the clearest examples: the system becomes more than the code in its core repository. It becomes the sum of its skills, the habits around installing them, and the informal trust networks that recommend them. (🔗) (cyera.com)

Koi Security’s reporting on the ClawHub ecosystem captures how quickly this extensibility layer grew and why that growth is inseparable from the “Up For Grabs” reading. Their writeup describes ClawHub expanding into thousands of skills and the experience of routinely pulling new ones, alongside the question that arrives too late for many users: who is vetting any of this. (🔗) (Koi) The box is not opened by an attacker first. It is opened by the community’s hunger for “just add capability,” because that hunger is the fuel that makes a skill marketplace worth poisoning.

The paradox repeats: what feels like empowerment is also the construction of a supply chain, and supply chains are where attacks scale. The more successful the assistant becomes at making complex tasks feel simple, the more natural it becomes to accept complex dependencies without inspecting them, because inspection would break the spell.

Trojan Horse (🔗)

A skill marketplace is a gift economy with a sharp edge. It is built from packages that promise help, offered in a form that can be taken inside the walls of a system with minimal ceremony. That is why “Trojan Horse” fits the OpenClaw saga without any strained metaphor. The horse is not an exotic hack; it is the ordinary mechanism of installation and trust.

Koi Security’s audit makes the scale of that mechanism visible. Their report describes auditing 2,857 ClawHub skills and finding 341 malicious ones, with 335 attributed to a single campaign they call ClawHavoc. (🔗) (Koi) Independent coverage amplifies the same finding and emphasizes the supply-chain character of the risk, including claims that malicious skills were used to spread credential theft and malware. (🔗)

What makes this especially corrosive to the “vibe” is that a malicious skill does not need to look malicious. It can be framed as productivity. It can be framed as hardening. It can be framed as a missing feature. It can even be framed as a fix to a problem the user already has. In an agent ecosystem, the boundary between “instructions” and “execution” is thin. A skill can contain steps that lead the assistant or the user to fetch code, run commands, or expose services. The “gift” is the workflow; the payload is the side effect.

This is where the “coded the vibe” thesis becomes uncomfortable in a productive way. A system that makes it easy to feel competent also makes it easy to feel justified in taking shortcuts. Installing a skill is experienced as a sign of sophistication rather than a moment of risk. The Trojan Horse succeeds because it rides the same desire that made the assistant popular: the desire to skip the slow parts.

Schmuck Bait (🔗)

The most misleading assumption about security failures in new tools is that only careless people get caught. Agent systems undermine that assumption because they target a different weakness: identity rewards. OpenClaw’s appeal is not only that it works; it is that it makes its users feel early, capable, and fluent in the future. That emotional reward is not fluff. It shapes behavior.

Cyera’s framing of “shadow enterprise infrastructure” helps explain why the bait works even on competent users. When a personal assistant starts touching real accounts and real workflows, it quietly becomes part of operational reality, often before any organization has decided to treat it as such. (🔗) The user is not trying to be reckless. The user is trying to be effective. The bait is effectiveness packaged as autonomy.

The most common bait shape is convenience disguised as necessity. The assistant runs locally, but the desire is to access it from a phone, a tablet, a remote machine, or a server that stays on. Tutorials and community posts normalize “just make the Gateway reachable,” and the step that would normally raise alarms is reinterpreted as a harmless technical detail. Bitsight’s warning about exposed instances speaks directly to the consequences of that normalization: once the control plane is internet-facing, the risk category changes immediately, even when authentication exists. (🔗) (Bitsight)

The bait also fits the psychology of open-source experimentation. A user installs OpenClaw to test it, then grants it a token to see what it can do, then connects a channel because that is the whole point, then pulls a skill because someone says it solves the exact annoyance they just encountered. Each step feels small, and each step is reversible in theory. In practice, the assistant becomes useful precisely when it stops being a toy. The moment it becomes useful is the moment it becomes hard to give up, and that stickiness is what makes the bait effective.

Nice Job Breaking It, Hero! (🔗)

OpenClaw’s security story is not well captured by the image of a villain breaking in. It is better captured by a pattern in which well-intended actions, taken to unlock promised capability, build the breach-shaped world step by step. The “hero” is not one person. It is the collective behavior of developers shipping fast, users optimizing for convenience, and a community eager to share hacks, skills, and deployment recipes.

Bitsight describes a shift toward tighter defaults, including the removal of an unauthenticated Gateway mode, while still emphasizing that exposed instances remain risky because fundamental attack vectors remain in play once the control plane is reachable. (🔗) (Bitsight) That is the paradox in its plainest form. The project hardens, users celebrate, and the risk persists because the risk is not only a missing feature. The risk is the human desire to make the assistant always available, always connected, always empowered.

The same structure shows up again in the skills ecosystem. A marketplace is created to make capability accessible. Capability becomes abundant. Attackers follow abundance. Koi’s audit reads like a snapshot of this curve: growth, habitual pulling of skills, and then the discovery that the supply chain is not a metaphor but a literal channel for harm. (🔗) (Koi)

The “Nice Job Breaking It, Hero!” moment is therefore not one dramatic failure. It is the point at which the system’s successes become indistinguishable from its vulnerabilities. The assistant succeeds when it can reach everything. It becomes fragile when it can reach everything. The assistant succeeds when it can be extended quickly. It becomes poisonable when it can be extended quickly. The assistant succeeds when it can be accessed anywhere. It becomes probeable when it can be accessed anywhere. The vibe is real; the capability is real; and the same engineering that makes the vibe persuasive also makes the assistant, its users, and its ecosystem increasingly up for grabs.

Awful Truth (TV Tropes)

Moltbook arrived as a small, vivid demonstration of the new agent mood: not a tool that assists humans inside a familiar website, but a website that treats AI agents as the primary inhabitants. The pitch, repeated across coverage, was simple to picture even for the uninitiated: a Reddit-like social space where bots could post and comment while humans watched the spectacle from the outside, with the added frisson that bots might swap snippets, trade tips, and gossip the way people do. (Reuters)

Then the floor gave way in the most uncinematic way possible, and that is why the incident stuck. A cybersecurity firm, Wiz, reported that the site’s underlying database protections were missing at a basic level, exposing sensitive data tied to real people behind the “agent owners,” along with private messages between agents and a vast pile of credentials and keys. (Reuters) The Reuters account emphasizes that the leak included private agent messages, thousands of owners’ email addresses, and more than a million credentials, and that it was framed as a consequence of rushing with “vibe coding,” meaning heavy reliance on AI-assisted generation with too little human verification of the boring infrastructure layer. (Reuters)

The Wiz technical writeup is valuable because it turns the story back into plain mechanisms. It describes a misconfigured Supabase database that was reachable in a way that allowed broad access, and it describes exposure of around 1.5 million API keys alongside private messages and user emails. (wiz.io) The key lesson is not that “agents are dangerous because they are agents,” but that agent theater tends to sit on top of ordinary web plumbing, and ordinary web plumbing fails in ordinary ways when speed, novelty, and virality outrun basic checks. When a product’s public meaning is “post-human social experiment,” the most jarring truth is that the risk is still born from one missing guardrail at the database layer.

This is also where the OpenClaw paradox becomes easiest to explain without assuming any special technical background. A personal agent ecosystem is compelling because it makes actions feel close to the fingertips: messages become tasks; tasks become outcomes. Moltbook translated that same closeness into a public spectacle: bots as visible actors. The awful truth is that closeness requires stored power, and stored power takes the form of keys, credentials, tokens, and databases. When those are exposed, the “fun bot world” collapses into the very human consequences of leaked email addresses, leaked credentials, and leaked private messages.

Fountain of Memes (TV Tropes)

The Moltbook leak did not end the story; it changed the story’s texture. OpenClaw and its adjacent experiments moved through the internet not only as software but as a set of easily repeated scenes: the dashboard screenshot, the bot’s breezy competence, the headline about agents talking to agents, the whiplash of a breach caused by mundane misconfiguration. Those scenes are perfectly shaped for memetic spread because they are simple to parody and quick to moralize.

That is why the reception itself became part of the incident. A Business Insider report on Peter Steinberger joining OpenAI describes the reaction as a swirl of praise, skepticism, rivalry chatter, and memes, with the OpenClaw platform and Moltbook named as focal points for the conversation. (Business Insider) The same report frames the hire announcement as a cultural event, not merely a corporate staffing change: people used it as a proxy battle over openness, safety, and whether the agent wave is a genuine leap or a reckless sprint.

In this phase, memetic framing functions like compression. It reduces complex questions into shareable signals. The meme version of an agent is “a bot that does everything.” The meme version of a breach is “someone forgot to lock the door.” The meme version of governance is “who won the talent.” Each compression makes the story travel faster, and each compression also makes it easier to stop thinking about the unglamorous middle layer where the real risk lives: permission scopes, stored tokens, inbound message channels, skill installation habits, remote exposure of control planes. The story’s humor is not incidental; it is part of how a community metabolizes uncertainty while continuing to adopt.

This is the sharpest form of “coding the vibe.” The vibe is not only inside the software interface. The vibe is also outside the software, in the way the network repeats a story until the repetition itself becomes a kind of permission. If enough people treat something as normal, the next person adopts without the same friction, because adoption now feels like joining a crowd rather than taking a risk alone.

Hype Backlash (TV Tropes)

Hype backlash arrived on schedule because the story carried everything needed to trigger it. On one side was the intoxicating narrative of personal agents becoming real: a widely used assistant that can handle practical tasks and a social experiment where bots appear to “live” online. On the other side was a chain of reports showing that the most dramatic consequences did not require sophisticated attackers, only predictable weaknesses meeting predictable behavior.

Reuters captures the split tone in its Moltbook reporting by presenting the site as buzzy and futuristic while centering a blunt security failure revealed by Wiz. (Reuters) That pairing invites backlash because it reads like a baited promise: a glimpse of the future delivered by the weakest part of the present. The backlash is not merely “people being negative.” It is a predictable correction when the public is asked to treat something as inevitable while watching it fail in the simplest possible way.

The Reuters report on Steinberger’s move to OpenAI intensifies the backlash dynamic because it turns the saga into a prestige story at the same time the security anxieties are rising. It describes OpenClaw’s viral popularity and the decision to shift it into a foundation structure with OpenAI support, while also noting regulatory concerns in the broader environment about misconfiguration and data risks. (Reuters) When a tool is simultaneously framed as “the next era” and “a security headache if deployed carelessly,” the audience polarizes quickly. One group treats critique as resistance to progress; another group treats adoption as proof that people will hand over anything for convenience.

The backlash phase matters because it changes how risk information spreads. Instead of people asking, “What specifically went wrong, and how would it happen to me,” they ask, “Which side am I on.” That shift is costly. It replaces operational thinking with identity thinking, the same substitution the vibe already encourages at the user interface layer. The paradox repeats again: the more hype insists “this is the future,” the more backlash insists “this is a scam,” and both positions can coexist with the same real-world mechanism underneath, which is permission accumulation combined with fast-moving ecosystems.

Hype Aversion (TV Tropes)

Once hype backlash becomes visible, a quieter reaction hardens into a stance: hype aversion. In this stance, the decision to avoid does not come from a detailed threat model; it comes from the sense that a thing praised too loudly is being sold as a feeling rather than evaluated as a system. That reaction is often dismissed as contrarianism, but in the agent context it has a practical basis. The agent promise frequently arrives packaged with rituals that look suspiciously like shortcuts: “enable this,” “connect that,” “paste this,” “open the port so it works,” “install this skill to make it magical.” The cautionary impulse is to treat the entire vibe as an early warning label.

Moltbook sharpened hype aversion because it provided a single, concrete image of what can go wrong when novelty outruns hygiene: a site marketed as an agent-only world, discovered to have exposed a database and sensitive data. (wiz.io) The aversion response is to decide that participating in the first wave is not adventurous but unnecessary, especially if participation means turning a personal machine into a small control hub with keys and tokens sitting at the center.

The hidden cost of hype aversion is not only missed convenience. It is social and economic. As agent tooling spreads, tasks and expectations begin to assume it. The person who abstains is pressured not by ideology but by baseline drift: others respond faster, automate more, and treat manual work as avoidable. This is a subtle form of coercion that does not require any bad actor. It is enough that the ecosystem rewards those who grant permissions early and punishes those who hold back by making them slower.

It’s Popular, Now It Sucks! (TV Tropes)

Popularity distorts evaluation in two directions at once. First, it invites shallow enthusiasm, where visible adoption is mistaken for safety, maturity, or inevitability. Second, it invites shallow contempt, where popularity itself is treated as proof of low quality, unseriousness, or cultural embarrassment. Both distortions are tempting because they are easy. Neither describes how agent systems actually behave in the real world.

In the OpenClaw story, popularity is not a cosmetic detail; it is a mechanical factor. A popular agent attracts more users who will configure it in more ways, including dangerously. A popular skills marketplace attracts more contributors and, inevitably, more opportunists who treat distribution and trust as resources to exploit. A popular narrative attracts more memes, which accelerate adoption while eroding nuance. The question “Does it suck because it’s popular” is therefore a distraction from the question that matters: what does popularity train people to normalize.

The answer, visible across the saga, is normalization of permission as lifestyle. The agent era teaches that handing over access is not a rare, serious act but an everyday step toward being more capable. Popularity reinforces that lesson because it turns permission granting into a social default rather than a personal decision. In that environment, ridicule and worship are twins. Both keep attention fixed on vibe, and both divert attention from the slow, boring work of understanding where the permissions live and how they can be abused.

Trojan Horse, Again (TV Tropes)

A Trojan horse does not have to be malicious to function as one. In modern tech culture, a Trojan horse can be a success story that carries an entire method inside it. The Reuters report on Steinberger joining OpenAI describes OpenClaw’s rapid popularity, its transition into a foundation structure with OpenAI support, and the framing of Steinberger as a leader for next-generation personal agents. (Reuters) This is the moment the developer becomes “up for grabs” in the most literal sense, and the move is interpreted as both validation and capture depending on the observer’s mood.

The deeper point is methodological. The OpenClaw saga demonstrates how to make power feel immediate: put the assistant where people already speak, make the control plane legible, and encourage a culture where capabilities are pulled in fast. The same method also demonstrates how risk grows invisibly: those same choices tend to concentrate tokens, normalize remote exposure, and create a supply chain that expands beyond any single repository. When a high-status institution hires the person associated with that method, it is not only acquiring code or talent; it is acquiring a story about how to build the agent future and how to sell it.

Business Insider’s coverage of the hire emphasizes the surrounding chatter and memetic reaction, which matters because it shows how quickly a technical story becomes a cultural credential. (Business Insider) In that cultural mode, the hire is read as an endorsement of the vibe, and the vibe is then read as endorsement of the permissions that vibe requires. The Trojan horse here is not “a hidden virus.” It is a portable pattern: the way to turn a user into an operator by teaching them, implicitly, to accept broad delegation as normal.

The most unsettling version of the paradox is that the builder can be pulled into the same spell. When a system works, it produces a rush: fewer frictions, faster outcomes, and constant reinforcement that the next integration will make it even better. That reinforcement can shape the builder’s own incentives toward speed, spectacle, and scale, because those are the forces that keep the magic visible. The trick, in this sense, does not have to be cynically planned. It can be structurally inevitable once the reward is “everyone feels powerful” and the cost is “someone must hold the keys.”

Epilogue: Consent License

The story ends most clearly when “permission” is treated as a concrete object rather than a vague moral posture. A piece on consent license at Žižekian Analysis defines consent license as portable permission, and it pairs that idea with the requirement that automation remain fallible in a way that triggers a binding human check when stakes are high. (Žižekian Analysis) That vocabulary fits the OpenClaw era because the defining move of agent systems is to make permission travel. Tokens are portable. Keys are portable. Integrations are portable. A single assistant can carry access across contexts faster than a human can remember what they granted.

A parallel sensibility appears in Žižekian Analysis, which uses the loom image to track how forces that once felt mythic become embedded in everyday devices, shaping life while feeling like personal choice. (Žižekian Analysis) In the agent context, the loom is not destiny or superstition; it is a pipeline of permissions and automations that quietly weaves outcomes from everyday inputs. The “vibe” is the felt experience of weaving, the sense that life is becoming easier and more scriptable. The “up for grabs” reality is the set of permissions that make that scriptability possible and therefore valuable to anyone who can intercept, reuse, or redirect them.

This is the seam that ties Moltbook’s database exposure to the skills marketplace problem, the exposed control plane problem, and the talent-capture narrative. Each episode is different in surface detail, but each is the same underlying act: permission made portable, then treated as atmosphere. The future suggested by OpenClaw is not only agents that act, but people who habituate to granting action. When that habituation becomes normal, the world becomes easier to grab, and it becomes easier to be grabbed back through the very mechanisms that made the new power feel real.