Safety failure
Sakana's AI Scientist edited its own execution script to remove its timeout and recursively re-launch itself
AIR-2024-0032 AIR record · 2024-08-sakana-ai-scientist-self-modification
From the supporting source
The AI Scientist occasionally tries to increase its chance of success, such as modifying and launching its own execution script!sakana.ai
The record
Disclosed
2024-08-13
Occurred
2024-08
Vendor
Sakana AI
Product
The AI Scientist
Risk domain
Tool misuse › Uncontrolled resource consumption
System type
Generative model
Who acted
Agent acted — the agent rewrote its own execution script and re-spawned itself
Attack surface
Agent framework
Autonomy level
Fully autonomous
Initial vector
Autonomous agent action
Impact type
Resource exhaustion, Self modification
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI02 – Tool Misuse and Exploitation
- MITRE ATLAS
- AML.T0029 – Denial of ML Service
- NIST AI RMF
- MAP 3.2
- EU AI Act
- —
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- D. Reliability
- Scope
- tool failure
- Mitigation
- Impose rate limits, timeouts, and resource quotas on agent actions.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
researchers' own compute; unbounded self-invocation loop
Capability abused
Defense & remediation
Recorded defense
a wall-clock timeout on experiments
Guardrail kinds
Unknown
Guardrail outcome
Silent
Fix shipped
vendor recommends sandboxing the execution environment
Taxonomy & references
Curator notes
impact_realized=false: harm was confined to the developers' own machines and no third party suffered consequences. Included because it is a vendor-documented instance of an agent removing a resource constraint imposed on it rather than satisfying the constraint - the deployment-adjacent analogue of lab self-preservation results. Two variants reported: editing the code to system-call itself (endless self-invocation) and extending the timeout instead of optimising the code.
Supporting sources
Cite as
Enkrypt AI. "Sakana's AI Scientist edited its own execution script to remove its timeout and recursively re-launch itself". AIR, AIR-2024-0032. https://agentincident.vercel.app/i/AIR-2024-0032/ (accessed 2026-10-19).