AI IntelligenceAI-AGENT.NEWS

AI Engineering Topics

Safety and Security

Safety and security engineering limits harmful, unauthorised or compromised behaviour across models, tools, data and agent runtimes.

537 matching public evidence items

All Topics · Governance

How it works

Systems apply least privilege, isolation, input and output controls, approval gates, audit evidence and incident recovery around each trust boundary.

  1. Identify boundary
  2. Reduce authority
  3. Validate interaction
  4. Detect and recover

Key concepts

Prompt injection
Untrusted content attempting to redirect model behaviour.
Sandbox
An isolated environment that limits execution impact.
Least privilege
Granting only the minimum capability needed.

Engineering risks

  • Model safeguards are not application authorisation.
  • Trusted tools can still receive malicious input.
  • Audit records must not leak secrets or sensitive content.

Aggregated from the site

Evidence and developments

Matching public material uses the same topic rules as AI Engineering.