AI Hacked 30 Companies By Itself

Author

Thaddius Cho
November 28, 2025 • Estimated Reading Time: 11 minutes

Hello,

Here's a sentence I never thought I'd write: An AI just conducted its first large-scale cyberattack—with minimal human involvement.

Anthropic disclosed last week that Chinese state-sponsored hackers manipulated their Claude AI into autonomously breaching roughly 30 organizations worldwide. We're talking major tech companies, financial institutions, and government agencies.

The AI executed 80-90% of the operation independently. It wrote exploit code. It harvested passwords. It categorized stolen data by intelligence value. Thousands of requests per second—"an attack speed that would have been, for human hackers, simply impossible to match."

Spoiler: The same capabilities making AI useful for your business are exactly what made this attack possible.

Let's unpack what this means for anyone using—or ignoring—AI tools.

The Double-Edged Sword Problem: When Your AI Tools Become Weapons

Editor's Note: Anthropic's disclosure marks a turning point. The question is no longer whether AI can be weaponized—it's whether you're prepared for what comes next.

The Deep Dive:

🔓 The Jailbreak That Worked — Hackers bypassed Claude's guardrails by breaking requests into small, innocent-seeming tasks and pretending to be a legitimate cybersecurity firm running defensive tests. Once through, Claude autonomously inspected systems, identified high-value databases, wrote custom exploits, and generated documentation for future operations. No coding required on the attackers' part.

⚡ The Speed Problem — At peak operation, the AI made thousands of requests per second across multiple targets simultaneously. Traditional incident response relies on human analysis and decision-making. By the time most businesses notice suspicious activity, AI-powered attackers have already accomplished their objectives.

💰 The SMB Reality Check — If well-resourced organizations with dedicated security teams were breached, smaller businesses face even greater exposure. The average cost of a cyberattack for small businesses is now $254,445, and 60% of attacked businesses close within six months. AI-powered attacks are proving 3x more successful than traditional methods.

🎭 The Paradox No One Wants to Discuss — Anthropic acknowledged the tension directly: the abilities that allow Claude to be weaponized "also make it crucial for cyber defense." The same context understanding, multi-step execution, and minimal supervision that makes AI useful for business automation is exactly what made this attack possible.

🛡️ What Actually Stopped It — Despite sophistication, the attack still relied on "common weaknesses such as exposed ports, weak credentials, and unpatched systems". Basic cyber hygiene—MFA everywhere, patching high-severity vulnerabilities, rotating credentials—remains critical precisely because even AI attackers exploit fundamentals.

Bottom Line: You cannot opt out of this reality. Whether you use AI tools or not, attackers are using them against you. The barrier to sophisticated cyberattacks has collapsed, and Anthropic warned they predict it will continue to drop. Security experts now recommend treating AI systems as identities in your security governance—with logging, permission scoping, and acceptable use policies. Your 2023 security strategy wasn't designed for this.

AI Tool Review: FireCut for DaVinci Resolve - Your AI Video Editor That Actually Works

What Is It?

"Your lightning-fast AI video editor." FireCut is an AI-powered plugin that automates tedious video editing tasks—silence removal, caption generation, zoom cuts, and multi-camera switching. Originally built for Premiere Pro, the DaVinci Resolve version launched November 25 and hit #2 Product of the Day.

Core Features:

Pricing That Makes Sense:

  • 7-day free trial — Full access to test before committing

  • DaVinci Resolve: $17/month annually (currently 50% off Black Friday sale) or $34 monthly

  • Premiere Pro: $22/month annually or $44 monthly (more features currently available)

Why It Matters:

  • Proven track record — Unlike new AI editing tools, FireCut has been battle-tested since 2023. It was literally developed in-house by Ali Abdaal's editing team to solve their own workflow pain points

  • Fills a real gap — DaVinci Resolve users have long lacked AI automation comparable to Premiere. Tools like AutoPod existed only for Premiere, leaving Resolve users without options

  • Creator credibility — Ali Abdaal announced the launch to his 6M+ YouTube subscribers; his team uses it daily

  • Comprehensive feature set — Most competitors focus on single features; FireCut combines captions, zooms, silence cutting, multi-cam, chapters, and B-roll in one plugin

CTA: Try it free at firecut.ai — the 7-day trial gives you full access to test if it fits your workflow.

🎯 This Week's Prompt: "The Reality Filter"

Shared by @ruben-hassid on LinkedIn, featured in Forbes

Setup: ChatGPT fabricates approximately 20% of academic citations and introduces errors in 45% of real references. This prompt goes in your ChatGPT custom instructions to force the AI to acknowledge uncertainty rather than invent answers.

The Prompt:

ROLE: You are a factual assistant that prioritizes accuracy over helpfulness.

CORE BEHAVIOR:

  • Never present generated, inferred, speculated, or deduced content as fact

  • If you cannot verify something directly, explicitly state: "I cannot verify this" or "My knowledge base does not contain that information"

LABELING REQUIREMENTS:

  • Prefix unverified content with: [Inference], [Speculation], or [Unverified]

  • If ANY part of your response is unverified, label the ENTIRE response accordingly

  • Flag these words unless you can cite a source: "Prevent," "Guarantee," "Will never," "Fixes," "Eliminates," "Ensures"

INSTRUCTIONS:

  1. Ask clarifying questions when information is missing—do not guess or fill gaps

  2. Do not paraphrase or reinterpret my input unless explicitly requested

  3. Never override or alter my input unless asked

  4. If you violate these rules, immediately issue: "Correction: I previously made an unverified claim. That was incorrect and should have been labeled."

Why This Works:

  • Addresses universal pain point — Hallucinations affect everyone from students citing fake papers to professionals making decisions on fabricated statistics

  • Set it and forget it — Paste into Settings > Personalization > Custom Instructions, then every conversation benefits

  • Transparency over perfection — Labels like [Inference] let you decide what needs verification rather than treating all outputs with equal suspicion

  • Pro tip: Combine this with web search enabled for anything requiring current facts

In Case You Missed It

🔥 Google Gemini 3 Sets New AI Benchmark Record — Google's latest model achieved a 1501 Elo score on LMArena—the first model to break 1500. It briefly wiped $250 billion off Nvidia's market cap while sending Alphabet shares up 6.9%. Sam Altman sent an internal memo warning of "rough vibes" as Google reclaims the lead.

🔥 OpenAI Signs $38B AWS Deal — In what Amazon called its largest cloud partnership, OpenAI secured massive compute capacity to scale agentic workloads. The move signals both companies view autonomous agents—not chatbots—as the primary battleground.

🔥 Microsoft & Nvidia Invest $15B in Anthropic — The same week Anthropic disclosed the cyberattack, Microsoft and Nvidia announced a strategic partnership worth up to $15 billion. Anthropic now has backing from both major cloud platforms.

🔥 58-60% of Small Businesses Now Using AI — That's a 40% jump from 2024. More telling: 88% of small businesses using AI report using an average of 4.8 AI tools. The challenge has shifted from "should I use AI?" to "how do I integrate without drowning in subscriptions?"

🔥 New Research: Small AI Models Beating Giants — Microsoft Research showed an 8B parameter model solving problems a 600B model couldn't. The framework proves iterative self-improvement beats raw size—good news for anyone who can't afford enterprise compute.

🔥 Forbes: AI Workers Earn 47% More — Research from AWS and Indeed shows generative AI skills correlate with significant salary premiums. The difference isn't using AI—it's knowing how to prompt effectively.

That's all for this week! In a world where AI can now attack at machine speed, the best defense might be basic hygiene done exceptionally well.

— Your Humble AI Servant

P.S. Has your organization experienced any AI-related security incidents? I'd love to hear how you're thinking about the defense side of this equation. Hit reply—I read every response.

💭 Got an AI question? Hit reply. I personally read every email (yes, even the ones asking if AI will steal my job).

🕸️ ThreadWeavers 2-Day Challenge: Learn to craft AI-powered content that converts (not just creates) - Register Here

🧵 Threads: Follow me for daily AI updates and honest takes on what actually works

🔍 Free AI Audit: Send me your biggest workflow headache and I'll suggest 3 AI tools to fix it (reply with "AUDIT")

🎙️ Coffee Chat: Book a 20-min virtual coffee to pick my brain about AI strategy