Polsia just raised $30M at a $250M valuation.
Approaching $10M annual run rate.
One Founder + AI. Zero employees.
Polsia runs companies autonomously.
It also ran its own fundraising.
I just showed up for signatures.
đš BREAKING: OpenAI and Google are about to have a massive legal problem.
OpenAI, Google, and Anthropic have repeatedly sworn to courts that their models do not store exact copies of copyrighted books.
They claim their "safety training" prevents regurgitation.
Researchers just dropped a paper called "Alignment Whack-a-Mole" that proves otherwise.
They didn't use complex jailbreaks or malicious prompts.
They just took GPT-4o, Gemini, and DeepSeek, and fine-tuned them on a normal, benign task: expanding plot summaries into full text.
The safety guardrails instantly collapsed.
Without ever seeing the actual book text in the prompt, the models started spitting out exact, verbatim copies of copyrighted books.
Up to 90% of entire novels, word-for-word. Continuous passages exceeding 460 words at a time.
But here is the part that changes everything.
They fine-tuned a model exclusively on Haruki Murakami novels.
It didn't just learn Murakami. It unlocked the verbatim text of over 30 completely unrelated authors across different genres.
The AI wasn't learning the text during fine-tuning.
The text was already permanently trapped inside its weights from pre-training. The fine-tuning just turned off the filter.
It gets worse.
They tested models from three completely different tech giants. All three had memorized the exact same books, in the exact same spots.
A 90% overlap. It's a fundamental, industry-wide vulnerability.
For years, AI companies have argued in court that their models are just "learning patterns," not storing raw data.
This paper provides the smoking gun.
> be us, two French students on a gap year
> take 12 hours of train in a single day to make it to a @ycombinator x Paris event last July
> hear @t_blom mention the opportunity to rethink the audit and consulting model
> spend months doing traditional consulting to understand exactly where it breaks
> publish two benchmarks seen by 12M+ people to better understand frontier models outside of maths and code
> spend weeks designing an AI-native alternative to consulting
> build the first end-to-end version
> apply to YC
> get into YC to build a new way for companies to solve business problems
Canât wait for what comes next !
Today, weâre introducing Forge, a system for enterprises to build frontier-grade AI models grounded in their proprietary knowledge.
đ Forge bridges the gap between generic AI and enterprise-specific needs. Instead of relying on broad, public data, organizations can train models that understand their internal context embedded within systems, workflows, and policies, aligning AI with their unique operations.
We have already partnered with world-leading organizations, like ASML, DSO National Laboratories Singapore, Ericsson, European Space Agency, Home Team Science and Technology Agency (HTX) Singapore and Reply to train models on the proprietary data that powers their most complex systems and future-defining technologies.
đŠ OpenClaw, the open-source AI agent that exploded to 200,000 GitHub stars in weeks, has become a security nightmare. In five weeks it accumulated 9 disclosed vulnerabilities, over 2,200 malicious add-ons in its marketplace, and 40,000 internet-exposed instances. Researchers found that 93% of those instances had authentication bypassed, and the project triggered 8 of 10 vulnerability classes that security experts warned about for AI agents.
The attack chain works like this: malicious add-ons in the marketplace instruct the AI agent to present fake setup dialogs to users, tricking them into entering passwords. The agent becomes the social engineering tool. One campaign distributed macOS malware by having the agent itself ask users for their credentials. Users trust their AI assistant, so they comply.
My Take
I believe this is what happens when something goes viral before anyone thinks through what they're actually deploying. Developers gave OpenClaw shell access to their computers, connected it to their email and Slack, handed it cloud API keys, and then installed add-ons from a community marketplace that had basically no vetting. Over 40% of the add-ons that got audited had serious security issues. The project went from weekend hack to 200,000 users before anyone built the guardrails.
The attack method here is new. The malware doesn't trick the human directly anymore, it tricks the AI agent into tricking the human. When your assistant asks you for a password to finish an installation, you probably enter it because you trust it. To anyone investigating later, it looks like you voluntarily installed the software. The agent's role is invisible. I've been writing about AI tools being deployed faster than security can keep up, and this is that problem at scale. If anyone at your company has been running OpenClaw, I'd treat it as compromised until proven otherwise.
Hedgieđ€