Experts warn of the 'first documented case of agentic ransomware' — dangerous JADEPUFFER attack run entirely by an LLM
THE SO WHAT
Agentic ransomware run entirely by an LLM—badly enough that it loses its own encrypted data—shows offensive automation is arriving before reliability. Treat LLM access, tool APIs, and prompt channels as part of your attack surface and start logging and rate-limiting them like you would any untrusted code execution.
READ THE SOURCE
MORE FROM THE WIRE
Applied AISpaceXAI releases Grok 4.5, which Elon describes as an ‘Opus-class model’
If Grok 4.5 is genuinely “Opus-class” but cheaper, the ceiling price for top-tier models just moved down. For teams building on proprietary LLMs, the negotiation leverage is shifting from brand to benchmarked capability-per-dollar.
Applied AIStill think you can spot AI? Here’s how to catch more convincing AI images, deepfakes and scams
If consumer guides are needed to spot AI images and scams, visual trust is now a UX and product requirement, not a niche security concern. Any product that relies on user-uploaded media should be budgeting for provenance, detection, and user education this year.
Applied AINew Meta AI tool lets users alter photos on public Instagram accounts
Letting an AI tool remix any public Instagram photo by default blurs the line between content and training data in a way users will feel directly. Brands and creators need to revisit what they post publicly — and how they communicate consent and boundaries around derivative AI use.
Applied AIOpenAI launches GPT-Live, a full-duplex voice upgrade that lets ChatGPT talk more like a person
Full-duplex GPT-Live turns ChatGPT from a turn-based tool into a real-time conversational surface. Any team building support, sales, or internal copilots now has to assume voice-native, interruptible agents are table stakes within 6–12 months.