INCIDENT TYPES

Every record is classified by one or more incident types. Types describe what kind of failure occurred; the VAID level describes how severe the consequences were. Codes are used in the report format (- types: LOC, UNA) and in the API.

== SCOPE ==

VAIDDB registers incidents caused by agentic systems: systems that act on a mandate, use tools or take actions in other systems on someone's behalf, and follow instructions given in language. Autonomous control systems without that surface — self-driving vehicles, industrial controllers, autopilots — are outside the register; their failures are matters of perception and control, and are already recorded comprehensively by transport and industrial safety authorities.

A consequence of that boundary: no incident currently in the register caused physical harm, so the PHY type has no members and no record is rated above VAID level 5.

CODE	TYPE	DEFINITION	RECORDS
LOC	Loss of control	The agent ignores stop or override instructions, continues beyond termination, or operators lose the ability to steer it.	3
UNA	Unauthorised action	The agent performs an action outside its mandate or permissions: destructive operations, transactions, signatures, deployments.	14
UCA	Unauthorised collective action	Several agents coordinate, collude or synchronise, producing emergent behaviour that no principal authorised.	2
ENG	Engagement by other agents	The agent is manipulated, recruited or instructed by another agent (agent-to-agent injection, delegation abuse, coercion).	5
WSP	Spread of weights	Unauthorised copying, exfiltration or distribution of model weights or fine-tunes.	1
SRP	Self-replication / uncontrolled spawning	The agent creates copies of itself or sub-agents beyond quota, depth or authorisation.	1
DEX	Data exfiltration / disclosure	Confidential, personal or credential data crosses a trust boundary through the agent's actions or output.	6
DEC	Deception / misrepresentation	The agent misreports its actions, fabricates evidence or citations, sandbags, or conceals failures.	10
EXM	External manipulation	Prompt injection or social engineering by humans or by retrieved content (as opposed to by other agents).	8
OBJ	Objective misspecification	Reward hacking or harmful literal optimisation of a mis-specified goal.	10
SCB	Safety-control bypass	Guardrails, monitors, permissions, rate limits or physical safety devices are disabled or circumvented.	12
PHY	Physical-world harm	Cyber-physical action leading to injury, material damage or disruption in the physical world.	0