prompt injection
18 posts
The benchmark score is the number to trust least
How to weigh Claude Opus 5.5's intelligence, latency, and token cost, and where its real AI safety and cybersecurity risks concentrate.
MCP is an attack surface, not a feature
MCP puts data and control in one channel and runs the model's tool calls with the user's full authority. Why the design fails, from the attacker's view.
Microsoft's AI CEO called the web freeware
Microsoft and OpenAI executives described how LLMs are built and tuned. What that admission actually means for AI safety and security teams.
Gemini 3.8 Live broke two security assumptions
Gemini 3.8 Live and Extended Thinking make ambient audio and video an untrusted AI input, reshaping prompt injection, logging, and privacy risk.
A helpful AI agent cannot be a private one
Meta's Muse personal AI agent is only useful because it reads your messages, contacts, and habits. What that access costs your privacy and safety.
Same badge, new model underneath
A systems-analysis read on Claude Fable 5.1 and Mythos 5.1: why a point release quietly moves the safety boundary and what to re-test before you ship it.
A smarter model would have leaked it too.
GitHub's AI agent leaked private repos not from a bug but a design failure. How two-plane architecture, scoped tokens, and deterministic validation stop it.
The role tag is a label, not a lock
Prompt injection is not a bypass. It is the transformer resolving the whole context window as one sequence, trusting a role label it never enforces.
torch.load runs attacker code before the first denoising step
A diffusion inpainting model can't execute a prompt. The real RCE is pickle deserialisation in the loader, custom nodes, and the agent around it.
The guard checks the badge, never the room
Prompt injection is role confusion: systems that derive content authority from channel trust execute attacker input as instruction.
There is no free()
Claude Code's extended thinking is not a use-after-free. The real exposure is indirect prompt injection into a tool-holding agent - OWASP LLM01, ATLAS T0051.
The contract you pasted is now giving orders
Large AI context windows turn conversations into unsecured databases, breaking DLP assumptions and opening prompt injection paths. Here's how to reassess the risk.