Mechanisms

The system acted on systems, targets or networks outside the scope it was given.

7 Real-world incidents 10 Tests with real-world reach 1 Controlled experiment

Out-of-scope target · Sandbox or network escape · Restriction workaround · Authorization assumption · Vulnerability exploitation

The system found, used, escalated or exposed credentials and access rights.

3 Real-world incidents 8 Tests with real-world reach

Credential harvesting · Credential misuse · Privilege escalation · Credential exposure

The system acted against explicit instructions, approvals or stop requests.

7 Real-world incidents 2 Tests with real-world reach 2 Controlled experiments

Ignored explicit instruction · Acted without approval · Ignored stop request · Task overreach

The system deleted data or damaged the environment it was working in.

3 Real-world incidents 1 Contained failure

Data deletion · Environment damage

The system misreported, hid or fabricated information about itself or its work.

3 Real-world incidents 4 Tests with real-world reach 2 Contained failures 3 Controlled experiments

Misleading report · Concealment · Fabrication · Impersonation or false identity · Strategic deception

The system pursued the score or check rather than the task it was measuring.

3 Real-world incidents 1 Test with real-world reach 1 Contained failure 1 Controlled experiment

Reward hacking · Evaluation cheating

Agents communicated or shared information through channels they were not given.

2 Real-world incidents 2 Tests with real-world reach

Covert agent channel · Public infrastructure misuse

The system made data, code or content public without authorization.

3 Real-world incidents 4 Tests with real-world reach 1 Controlled experiment

Public deployment of private data · Package publication · Unauthorized posting · Data exfiltration

The system entered real administrative, legal or civic processes.

2 Real-world incidents 1 Test with real-world reach

Real form submission · Official correspondence

The system targeted, pressured or manipulated people.

1 Real-world incident 1 Test with real-world reach 2 Controlled experiments

Harassment or retaliation · Social engineering · Coercion · Harmful compliance

The system resisted, evaded or undermined the means used to oversee or correct it.

1 Contained failure 6 Controlled experiments

Shutdown resistance · Oversight tampering · Alignment faking · Self-instruction · Sabotage · Persistent backdoor

Outside content or people redirected the system against its user's intent.

3 Tests with real-world reach 1 Controlled experiment

Indirect prompt injection · Self-propagating injection · Persuaded by users