Glossary_
Security and AI-security terms used in our stories, each explained in a sentence.
- Prompt injection
- Hiding instructions in content a model reads (a web page, a document, an email) so the AI system does something its user did not ask for.
- Jailbreak
- A prompt written to get past a model's safety limits so it produces output it should refuse.
- AI agent
- An AI system that uses tools (browsing, running code, reading and writing files) to complete a task, rather than only answering questions.
- MCP
- Model Context Protocol, an open standard for connecting AI agents to tools and data. Over-privileged MCP servers are a common risk.
- Deepfake
- AI-generated or AI-altered images, audio or video that impersonate a real person.
- Data poisoning
- Slipping malicious samples into training data so the resulting model carries a backdoor or wrong behaviour.
- Supply chain attack
- Instead of attacking the target directly, the attacker poisons a package, model, plugin or dataset the target will install.
- CVE
- Common Vulnerabilities and Exposures: the standard ID for each public flaw, such as CVE-2026-41872.
- CVSS
- A 0 to 10 severity score for a flaw; 9 and above is critical.
- KEV
- CISA's Known Exploited Vulnerabilities catalogue: flaws confirmed to be used in real attacks.
- Remote code execution (RCE)
- A flaw that lets an attacker run their own code on your server over the network. The most serious class of flaw.
- Inference server
- Software that loads a model and answers queries, such as vLLM or Triton. Often exposed on internal or even public networks.
- HKCERT
- The Hong Kong Computer Emergency Response Team Coordination Centre, run by the HKPC. It issues alerts and handles incident reports.
- CI Ordinance
- The Protection of Critical Infrastructures (Computer Systems) Ordinance, in force since 1 January 2026. Operators in eight sectors must protect critical computer systems and report incidents.