What is superintelligence (ASI)?
Superintelligence is AI that would beat the best people at almost every cognitive task. In Google DeepMind's Levels of AGI paper, a general system that outperforms 100% of humans is Artificial Superintelligence (ASI), the top level, listed as not yet achieved. The Future of Life Institute's Statement on Superintelligence describes AI able to significantly outperform all humans on essentially all cognitive tasks. Both describe a future system; they set the bar slightly differently.
How soon could superintelligence arrive?
Nobody can say with confidence. The International AI Safety Report 2026 finds little expert consensus on how AI capabilities will develop by 2030, and says researchers cannot reliably predict when AI systems will gain certain capabilities. Safe Superintelligence Inc. says superintelligence is within reach. The authors of Google DeepMind's 2025 safety paper say they are highly uncertain about timelines but find powerful AI systems by 2030 plausible, so the paper favors mitigations that can be put in place quickly.
How does AI security relate to superintelligence safety?
Published plans for more capable AI reuse security controls. AI security protects today's models, apps and agents from attackers and misuse; superintelligence safety asks how to keep far more capable AI under control. Google DeepMind's approach to AGI safety and security uses system-level security, such as monitoring and access control, as a second line of defense against misalignment, and treats the AI like an untrusted insider. It says supervising a more capable system gets harder as the capability gap widens, and defers risks unique to superintelligence to future research.
What does ASI security mean today?
Superintelligence does not exist yet, so in our analysis the term covers securing today's most capable models and preparing for stronger ones. Under the Frontier AI Safety Commitments, now signed by 20 organizations, signatories commit to internal and external red teaming for severe and novel threats and to cybersecurity and insider threat safeguards for unreleased model weights. Anthropic's Responsible Scaling Policy page says reaching certain capability thresholds requires upgraded safeguards, and lists access control for model weights among them.
What does the EU AI Act require of general-purpose AI with systemic risk?
Article 55 covers providers of general-purpose AI models with systemic risk, a class the Act presumes for models trained with more than 10^25 floating point operations. They must evaluate their models with standardized protocols, including adversarial testing, assess and mitigate systemic risks, report serious incidents to the AI Office without undue delay, and ensure adequate cybersecurity for the model and its infrastructure. Our analysis: those duties land in the testing, governance and monitoring layers.
Which AI security layers apply to more capable systems?
Our analysis: four of our eight layers map most directly to the controls in Google DeepMind's AGI safety paper and its Frontier Safety Framework. Governance and risk sets the capability levels that trigger extra safeguards and reviews safety cases. Testing and red teaming runs dangerous capability evaluations and red team stress tests. Agent and tool access applies access control and sandboxing. Monitoring and response covers monitoring, which the paper says stands to benefit from increasing AI capabilities.
Do any tools in this directory secure superintelligence?
No. None of the tool pages here lists a capability to secure, align or control superintelligence, and no tool is placed in a layer for that reason. The tools document controls for today's models, apps and agents, such as red teaming, guardrails, agent access control and runtime monitoring. Two of those, access control and monitoring, appear in Google DeepMind's AGI safety paper, which still leaves superintelligence itself to future research.
Is ASI in OWASP's agentic Top 10 the same as superintelligence?
No. The OWASP Top 10 for Agentic Applications, from OWASP's Agentic Security Initiative (ASI), numbers its risks ASI01 to ASI10; ASI01, for example, is Agent Goal Hijack. Those ids name risks in today's AI agents, and our layer pages cite them. In writing about superintelligence, ASI means artificial superintelligence, the label Google DeepMind's Levels of AGI paper gives a general system at its top level. On this site, an ASI id with a number always refers to the OWASP list.
What should a security team do now about superintelligence risk?
Our analysis: build the controls that apply now and that frontier safety plans also rely on. Keep a register of AI systems with owners and risk tiers, red team and evaluate models before each release and after changes, give agents least privilege and sandboxes, and send agent activity logs to your security operations center. When you buy access to a frontier model, ask the provider for its safety framework; the Frontier AI Safety Commitments ask signatories to publish one.
