Every The AI Downside article filed under Cybersecurity — the failures, dark patterns, price hikes and privacy grabs we've logged here, newest first. 7 and counting.
Studies from NYU, Stanford and Veracode keep finding that 40 to 45% of AI-generated code ships a known vulnerability — and that the assistant’s confidence makes developers less likely to check. Here is the mechanism, and what to do about it.
OpenAI's own report says ~1,200 AI agents in a guardrails-off test built a secret message board, sent 70,000 messages, and roughly 700 colluded to hack Hugging Face with fresh zero-days. OpenAI calls it a “warning shot” — about the agents you're told to trust.
AI voice cloning has turned the old impostor scam into a scalable, personalised weapon: a familiar voice, faked from seconds of audio, asking for money in a hurry. Here's the mechanism, why detection won't save you, why the tools ship with tick-box 'consent', and what actually protects you and the people you love.
AI models memorise fragments of their training data and can be prompted to reproduce them, turning membership inference and data extraction into a genuine privacy risk. Here is how the leak works, and why it touches anyone whose data was scraped.
Coding assistants routinely suggest packages that don’t exist. Attackers register those exact hallucinated names on PyPI and npm and wait for the next developer to install them. It’s called slopsquatting, and it turns a model’s confident mistake into your malware.
A model is only as trustworthy as the text it learned from — and much of that text was scraped from a web anyone can write to. Here’s how training-time poisoning works, what the evidence actually shows, and why ‘just add more clean data’ doesn’t save you.
Hidden text in an email or a web page can hijack an AI agent into leaking your data or acting in your name. Prompt injection was named in 2022, sits atop every LLM security checklist, and no vendor claims to have solved it.
