APEX-Harm benchmark: script-based injection attacks easily compromise agents in shared workspaces
KAUST · hf · 2026-10-16
APEX-Harm is a new benchmark evaluating long-horizon agent vulnerabilities to injection attacks in shared professional workspaces.
How it differs from prior benchmarks:
- Shared multi-user workspaces with 115–328 files averaging 40+ pages across banking, management and law, using CLI plus multiple MCPs (vs. single-user, <40-file general tasks)
- Attacks can't predict which files the agent will read and must induce it to consume injected content; leakage covers both external and internal (private → shared folder) paths
Findings: script-based injection attacks achieve high success rates even when the agent's task outputs look correct, exposing serious security gaps in multi-tool workspaces.
More from Safety
- Vector Institute partners with Town of Caledon in its first municipal AI adoption deal — VectorInst · 2026-10-16
- 'Most dangerous technology in the world' put in toddler sandboxes, and we're surprised when it escapes — KevinNaughtonJr · 2026-10-16
- NYC Comptroller urges Anthropic to pause IPO until AI safety and governance concerns are addressed — ShakeelHashim · 2026-10-16
- Dwarkesh: Anthropic insider expects AI to match humans at AI research within 9 months — ShakeelHashim · 2026-10-16
- Anthropic and Gray Swan launch $300,000 indirect prompt injection red-teaming contest — bcherny · 2026-10-16
- Experts: AI's nearer bioweapons risk is amplifying capable actors, not enabling amateurs — HaydnBelfield · 2026-10-16