Prompt Hacking Mexico; Perplexity vs OpenClaw; Fashion Backlash
Today's AI Outlook: ☀️
Claude Used To Breach Mexican Systems
Attackers reportedly used Anthropic’s Claude to help breach Mexican government systems, exfiltrating roughly 150GB of sensitive data. According to reports, the haul includes taxpayer files, voter records and internal credentials, with activity spanning several weeks before accounts tied to the prompts were banned.
This was not an AI going rogue. It was humans using a frontier model to accelerate reconnaissance, script generation and iteration. Mexico’s tax system reportedly covers 195M records, and credential reuse across agencies may have widened the blast radius. Officials are still assessing the full scope.
Why it matters
The barrier to entry for offensive cyber operations is collapsing. Tasks that once required layered expertise can now be scaffolded through iterative prompting. AI does not replace attackers. It compresses the learning curve and speeds up execution.
Institutional defenses, meanwhile, still move at institutional speed.
The Deets
- Data allegedly includes taxpayer records, voter rolls and credentials
- Activity persisted for weeks before detection
- Accounts associated with misuse were banned
- Scope and exposure still under investigation
Key takeaway
AI may not be the most sophisticated hacker. It may be the most accessible one. The asymmetry between offensive acceleration and defensive modernization is widening.
🧩 Jargon Buster - Reconnaissance: The early phase of a cyberattack where attackers gather information about systems, credentials and vulnerabilities before exploiting them.
💻 Power Plays
Perplexity Launches 19-Model AI “Computer”
Perplexity AI just introduced Perplexity Computer, a hosted agent system that orchestrates 19 different AI models across tasks. The $200 per month Max tier offers a 10,000 credit bank, with consumption-based pricing and optional manual model selection. Look out OpenClaw?
What’s happening
Users describe an outcome. The system spins up sub-agents that can browse, code, connect to apps and execute workflows inside sandboxed environments. CEO Aravind Srinivas framed it as a rebuttal to single-model ecosystems, taking a public swipe at Anthropic’s closed collaboration loop.
Each task runs in its own environment, and the company claims agents can persist for months.
Why it matters
Multi-model flexibility has existed in creative tools. Wiring it directly into an autonomous agent platform shifts the stack. If orchestration becomes the product, model vendors risk commoditization.
At the same time, shared infrastructure introduces shared risk. When memory, logs and credentials sit in vendor-controlled environments, isolation is a design promise, not a physical boundary.
The Deets
- 19 models orchestrated per workflow
- Hosted cloud sandbox per task
- $200 per month Max tier with 10K credits
- Manual model selection optional
Key takeaway
Execution is becoming modular. The company that best coordinates intelligence layers may matter more than the one that builds them.
🧩 Jargon Buster - Multi-model orchestration: A system that dynamically routes subtasks to different AI models based on strengths like reasoning, coding or vision.
Gemini Expands Into Persistent Mobile Galaxy
Google and Samsung unveiled deeper Gemini integrations across the new Galaxy S26 lineup, positioning it as a multistep execution agent that reads chats, opens apps and stages transactions.
What’s happening
Gemini is no longer framed as reactive Q&A. It is pitched as a context-aware assistant that operates across apps, memory and workflows. Samsung’s devices also support swappable agents, including Perplexity.
The smartphone is turning into a local agent hub.
Why it matters
The race is no longer about smarter responses. It is about persistent presence. Cloud-native agents like OpenClaw run continuously with structured files and cross-tool autonomy. Handset AI is attempting to simulate permanence within OS constraints.
This is the convergence of personal compute and agentic autonomy.
The Deets
- Multi-app execution inside mobile OS
- Swappable AI agents on Galaxy S26
- Expanded cross-session memory and task staging
Key takeaway
If persistence defines the category, mobile AI must compete with cloud agents that were built persistent from day one.
🧩 Jargon Buster - Persistent agent: An AI system that maintains memory and context over long periods, allowing continuous task execution instead of one-off responses.
🤖 Research & Models
Robotics And The "Shared Brain"
Physical Intelligence released deployment data for its π0.6 foundation model acting as a shared brain for third-party robots.
What’s happening
Instead of building custom AI stacks, robotics firms plug into a shared foundation model, fine-tune on proprietary task data and run locally. Early deployments showed laundry-folding interventions dropping 50% and warehouse throughput rising.
Meanwhile, Alphabet’s Intrinsic is integrating more tightly with Google’s AI infrastructure, aligning industrial robotics with frontier models like Gemini.
Why it matters
Robotics is decoupling hardware from intelligence. Swappable AI “brains” mirror what the Air Force just demonstrated midair.
In a separate test, the U.S. Air Force flew Anduril Industries’ YFQ-44A drone and swapped its AI brain mid-flight, separating flight control from combat decision logic. Software updates no longer require landing and rewiring.
The Deets
- π0.6 model shared across vendors
- 50% drop in manual interventions in pilots
- Air Force AI swap midair demonstration
- Intrinsic aligning closer with Google
Key takeaway
AI is becoming modular infrastructure. In software and robotics alike, intelligence is shifting from vertically integrated to API-first and swappable.
🧩 Jargon Buster - Foundation model for robotics: A large, general-purpose AI model trained on diverse physical tasks that can be fine-tuned for specific robot behaviors.
🎭 Culture & Commerce
Gucci Faces Backlash Over AI Ads
Gucci released AI-generated visuals promoting its Primavera campaign ahead of Milan Fashion Week and quickly faced backlash.
What’s happening
The campaign disclosed synthetic imagery alongside traditional shots. Critics called the visuals cheap and misaligned with a $11.6B luxury house built on craftsmanship. Gucci has experimented with AI before, including synthetic runway clips and NFT collaborations.
Why it matters
AI adoption in creative industries is inevitable. Quality bar and brand alignment remain decisive. When synthetic output undercuts premium positioning, the tech becomes the story instead of the product.
The Deets
- AI images tagged with disclosure
- Social backlash and boycott threats
- Prior AI runway and NFT experiments
Key takeaway
In luxury, perception is product. AI must elevate craft, not dilute it.
🧩 Jargon Buster - Synthetic media: Content generated by AI systems rather than captured through traditional photography or video production.
⚡ Quick Hits
- Anthropic retired Claude Opus 3 but gave it “Claude’s Corner,” a weekly AI-written newsletter while keeping API access for paid users.
- MatX raised over $500M led by Jane Street and Situational Awareness to build next-gen AI chips.
- Adobe added Quick Cut to Firefly, turning raw or AI-generated footage into rough edits from text prompts.
- DeepSeek reportedly withheld V4 access from Nvidia and AMD, favoring domestic firms.
- NTT DOCOMO and Keio University improved 5G robot teleoperation, cutting jitter and boosting motion smoothness by 59%.
🧰 Tooling Around
- Perplexity Computer: A 19-model orchestration agent designed for long-running autonomous workflows. Consumption-based pricing with manual model control.
- Quick Cut by Adobe Firefly: Converts raw footage into a structured first edit from a prompt. Aimed at speeding up creative iteration.
- Comet by Perplexity: Turns bookmarked content into structured research workflows and scheduled discovery tasks.
- Bland AI: Voice cloning platform that automated Soulja Boy’s voice, generating 30M views in 24 hours and driving enterprise signups.
Today’s Sources: AI Secret, The Rundown AI, Robotics Herald