#llm
7 posts
Reading DeepSeek official charts: what V4.1 Flash-1017 actually improves
I pulled the benchmark table, the performance-cost chart and the case images from DeepSeek's official changelog. The update is about doing work, not about being smarter — and the price did not change.
873 bugs, 54 models: notes on the 2026 LLM security report
2,467 white hats tested 54 LLM and agent products from 25 vendors and found 873 vulnerabilities, 608 of them LLM-specific. Prompt injection still ranks first; a hardcoded API key exposed keys for 239 models at once.
An agent identity is more dangerous than its capability
In the LLM security report, agent identity and permission abuse plus goal hijacking account for two categories. Delegation chains, memory defects and forged tool responses can give a low-privilege agent high-privilege credentials.
Prompt injection is still number one, because we still treat data as instructions
In the 2026 LLM security report, prompt injection remains the most common finding. It is hard to fix not because models are dumb, but because systems lack a boundary between data and instructions.
PwnEscape and agent sandboxes: a microVM is not an absolute boundary
PwnEscape exploits an L2/L3 offset mismatch in virtio-net/TAP to reach the KVM host from a guest, tracked as CVE-2026-99999. It is a reminder that isolating untrusted code in a microVM assumes the host kernel has no reachable path from the guest.
239 model keys behind one hardcoded variable
The sharpest example in the LLM security report: a product hardcoded cloud API keys for 239 models in its frontend. Key leaks are not new, but LLMs amplify them — one key is a whole fleet of models and a bill.
GEO manipulation: when model answers are decided by search ranking
When DeepSeek opened its search API it mentioned filtering SEO spam, GEO manipulation and low-quality AIGC. GEO optimizes the probability of being cited by a model, not a page rank. Here is how it works and why it is hard to defend.