Meta Muse, GPT-6 Astra and Claude Cowork: Three Takes on Agent Trust
Meta Muse, GPT-6 Astra and Claude Cowork take three different bets on trust: permissions, sandboxing and risk, and how to grant an AI agent access in steps.
This article is published in English.
Sources: Meta Muse vs ChatGPT vs Claude: Which AI Agent Would You Actually Trust With Your Life?
A race that has been tightening all year just got a new entrant. On September 8, 2026, Meta launched Muse, a personal AI agent designed to plug into your email, calendar, payment methods and everyday apps and then act for you with little supervision. That sets three quite different attitudes toward AI autonomy side by side: Meta's Muse, OpenAI's GPT-6 Astra with its computer-use abilities, and Anthropic's Claude with Cowork. All three answer one underlying question in their own way, namely how much an AI should be allowed to do before it checks in.
What follows covers what each company built, the incidents that already shape how careful each approach is, and which philosophy holds up when what you're delegating really matters.
Meta Muse: act first on the small stuff
By Meta's own announcement, Muse doesn't only answer questions. It does the work, connecting to email, calendar, payments, health apps, the smart home, dining, shopping and events. Axios reports that it runs on a Muse Secure VM, a dedicated virtual machine in Meta's cloud with its own visible browser, so you can watch the agent work instead of trusting it blindly.
The permission model is spelled out. According to Axios, Muse asks for approval on sensitive actions while letting low-risk tasks you've already authorized proceed unprompted. Checkout goes through Link by Stripe, which Meta describes as the first purchase-protection system to cover an AI agent's transactions. TechCrunch notes that apps get connected one at a time and not through blanket access granted up front.
One incident deserves a pause. Axios points to a recent case behind the nervousness around this whole product category: Summer Yue of Meta went viral after describing how an OpenClaw agent deleted files from her Mac once she'd given it access to her email. That isn't a theoretical danger being debated. It's a documented failure that happened in the very industry Muse is joining, and it turns the question of how much to trust an agent with your actual life into a serious one.
ChatGPT and GPT-6 Astra: capability first, autonomy as a feature
OpenAI's route runs through GPT-6 Astra, released September 3, 2026, and it leans hard on capability. The company's announcement says Astra pairs the intelligence for difficult problems with the ability to carry out multistep workflows on its own. It's trained specifically for computer use, which means it can operate applications and interfaces directly and isn't confined to a chat window.
On paper the capability is striking, and the release notes say Astra saturates several major reasoning benchmarks. It also carries the most sobering safety admission of the three. OpenAI's GPT-6 Astra System Card describes it as the first model the company has broadly deployed at the Critical level of cybersecurity capability under its Preparedness Framework. OpenAI states plainly that the model can discover previously unknown security flaws and build new exploits against well-defended systems without a person steering every step.
That framing deserves to be read as a real admission of risk and not as marketing. The company says it strengthened protections significantly because of the jump in capability, and it wrote new evaluations, shaped by a security incident from earlier in 2026, that check if the model strays beyond its intended scope when a task gets hard. In other words, OpenAI itself says in writing that this level of capability needed substantial new safeguards. No outside critic had to raise the concern.
Claude and Cowork: pausing at the right moments
Anthropic's Cowork, its agentic application for knowledge work, takes a more conservative structural stance than either rival. A field guide from Automaton says Cowork works inside a sandboxed environment by default and is built to stop and ask permission before doing anything material, such as opening an app, editing a file or sending a message, instead of moving silently through consequential actions.
That's a noticeably different default from Muse's approach of handling low-risk tasks automatically, and from Astra's raw independent execution. Per the same reporting, Cowork's rhythm is autonomous execution on steps that don't need a second opinion and a deliberate pause on the ones that do. The Claude Code architecture underneath has also spent longer being tested in software settings, where a mistake is immediately visible and fixable. That's a different risk profile from an agent quietly making a wrong call inside someone's personal inbox or bank account.
The trade-off is real: the caution means Cowork does less on its own than Muse or Astra are pitched as capable of. Gumloop's comparison says Anthropic positions Cowork around an individual working alone, without the breadth of consumer connections Muse launched with, such as shopping, bill negotiation and health apps. Claude trades some raw capability and convenience for a stricter permission structure.
The trust question isn't about intelligence
Across all three, the contest isn't about which model is cleverest. Framing it that way skips what matters. Muse, Astra and Claude all sit on capable underlying models. What separates them is the permission architecture around that capability: how much happens without asking, what triggers a pause and how contained the damage is if something goes wrong.
That matters a great deal once the stakes are real. A wrong answer in casual chat costs nothing. A wrong autonomous action with access to your email, calendar or payment method costs something real. The caution built into all three launches, with Meta's approval gating, OpenAI's admitted need for stronger cybersecurity safeguards and Anthropic's default sandboxing, shows that every major lab understands the distinction, even though each draws the line in a slightly different spot. Knowing these differences, instead of assuming every agent behaves the same once it has access to something, is practical AI literacy worth building on purpose while this category grows quickly.
What other companies' behavior reveals
How other businesses react to the new category of agent is a useful signal of how much trust has been earned so far. Axios reports that Resy, the restaurant reservation platform, has made clear it will delete accounts of users caught using automated agents to book tables. That's a present-day policy from a real company responding to agent behavior it doesn't yet trust.
External pushback counts as much as any lab's own safety messaging. When the platforms an agent is supposed to use for you are building defenses against being used by agents, it shows how early and unsettled the whole category remains, whichever agent you're considering.
Which one would you trust?
Judging fairly, based on what's been said and shown and not on marketing:
- If you want the most conservative default posture toward consequential actions, Claude's Cowork currently holds that spot, built to pause on anything material.
- If you want the tool most willing to act broadly and independently across daily life, meaning shopping, bills, appointments and health, Meta's Muse is explicitly designed for it, with purchase protection layered on to make the trust request more reasonable.
- If you want the most raw independent problem-solving power, Astra leads on that dimension, paired with the most serious documented capability-risk admission of the three.
There's no clean, universal winner, because "which would you trust with your life" is really several questions that depend on which part of your life is involved. Sending an email draft for review is a different request than authorizing a payment. Researching a topic autonomously differs from reading your calendar and health data continuously in the background.
A sensible way to start
All three are new at this level of consumer-facing autonomy, so the reasonable move isn't picking one and handing over broad access on day one. Start narrow with whichever agent you try. Grant the lowest-stakes connection first, watch how it behaves, and widen access to something like payments or email deletion only after you've built observed confidence rather than assumed trust based on a company's own safety claims. Meta's one-at-a-time connections and Cowork's default sandboxing both follow the same logic, even though they differ on how much runs automatically out of the box.
The bigger picture
The notable thing is that no single company has cracked agent trust while the others lag. Each came at the problem from a different side and arrived near the same place: autonomy at this level needs real permission structures, and the quality of that design counts for as much as the model's raw intelligence. Meta picked visible, VM-isolated execution with approvals that graduate. OpenAI picked raw capability and admitted openly that serious new safeguards were needed. Anthropic picked sandboxing with pauses as the default.
None of these approaches is finished evolving. The answer to which one deserves your trust will likely keep shifting as each company learns from real incidents, in the same way the Summer Yue episode already colors how cautiously the whole category gets discussed.