> ## Content Index > Fetch the complete content index at: https://www.aigl.blog/llms.txt > Use this file to discover other available public pages before exploring further. # Frontier Safety Framework - URL: https://www.aigl.blog/frontier-safety-framework/ - Published: 2026-09-19T08:03:20.000Z - Updated: 2026-09-19T08:03:20.000Z - Description: Google DeepMind's Frontier Safety Framework v1.0 sets out protocols for detecting Critical Capability Levels in Autonomy, Biosecurity, Cybersecurity and ML R&D, and pairs each risk domain with graded security and deployment mitigations. - Author: Jakub Szarmach - Tags: Framework, AI Safety, Risk Management, AI Security, Agentic AI, AI Governance, #aigl-library ## ⚡ Quick Summary Published by Google DeepMind, this framework sets out an initial set of protocols for addressing severe risks that may arise from the powerful capabilities of future foundation models, and is intended to complement Google's existing AI responsibility and safety practices. It introduces Critical Capability Levels (CCLs): capability thresholds at which, absent mitigation, models may pose heightened risk. The first set covers four domains, with CCLs such as Autonomy level 1, Bio amateur enablement level 1, Bio expert enablement level 1, Cyber autonomy level 1, Cyber enablement level 1, and Machine Learning R&D levels 1 and 2. The document describes how models are tested using "early warning evaluations" every 6x in effective compute and for every 3 months of fine-tuning progress, and how a response plan is formulated when evaluation thresholds are met. Two mitigation categories are defined: security mitigations to prevent the exfiltration of model weights (Levels 0-4, mapped approximately to RAND L3-L5) and deployment mitigations to manage access to critical capabilities (Levels 0-3). The framework states that it is exploratory, based on preliminary research, informed by Responsible Capability Scaling, and aimed for implementation by early 2025. ## 🧩 What’s Covered - **Framework**: defines the three central components — Critical Capability Levels, periodic evaluation of frontier models, and the application of mitigations — and states that the framework is built around capability thresholds; Figure 1 depicts how the components relate. - **Evaluating frontier models**: describes "early warning evaluations" with a specific "pass" condition, a cadence of every 6x in effective compute and every 3 months of fine-tuning progress, and a safety buffer so evaluations are passed before a model is 6x or 3 months away from a CCL; effective compute is defined as a scaling-law metric integrating model size, dataset size, algorithmic progress and compute. - **Applying mitigations**: explains that reaching evaluation thresholds triggers a response plan based on CCL analysis, evaluation results, additional flagged risks and deployment context; if appropriate mitigations are not ready, further deployment or development is put on hold. - **Security Mitigations**: a five-level table (0-4) for preventing weight exfiltration, from industry-standard controls through controlled access, lockdown of unilateral access, high-trust developer environments and advanced protection, with approximate mappings to RAND L3 to L5 and measures such as SLSA Build L3 provenance and confidential-compute TPUs. - **Deployment Mitigations**: a four-level table (0-3) covering safety fine-tuning and misuse filters, mitigations targeting the critical capability, a safety case with red-team validation and a robustness target, and prevention of access — the last noted as an open research problem. - **Critical Capability Levels**: a table listing each CCL by risk domain with its rationale, including autonomous resource acquisition, amateur- and expert-level biothreat enablement, opportunistic and severe cyberattack automation, and ML R&D acceleration (e.g. 3X algorithmic progress) and full automation of the AI R&D pipeline. - **Future work**: priorities including greater precision in risk modelling, capability elicitation, mapping CCLs to mitigation levels, additional risk domains (misaligned AI, chemical, radiological and nuclear, higher CCLs) and involving external authorities and third parties. - **Acknowledgements**: names the staff who developed the framework and those who contributed, and credits METR for expertise on Responsible Capability Scaling. ## 💡 Why it matters? For teams building, assessing or governing frontier models, the document shows one concrete way to turn abstract risk claims into machinery: a capability threshold is defined, an early warning evaluation is designed to flag it in advance, and specific security and deployment levels are attached to the response. Security teams can compare the weight-protection levels against the RAND and SLSA benchmarks the framework names. Policy and assurance staff can use its structure when reading comparable responsible-scaling policies, which the document itself links to. ## ❓ What’s Missing The document does not print a publication date or reference number, and the CCLs are explicitly preliminary. Pass conditions, thresholds and the designs of the early warning evaluations are not given, nor is the mapping from individual CCLs to specific security and deployment levels, which is deferred to later versions. Deployment Level 3 is described as an open research problem. Misaligned AI, chemical, radiological and nuclear risks and higher CCLs are named as possible future inclusions rather than covered, and external alerting, third-party involvement and post-mitigation outcomes are still being explored. ## 👥 Best For Safety and policy leads at frontier model developers designing capability-threshold frameworks; security engineers implementing model-weight protection controls; governance and assurance staff comparing responsible-scaling approaches across labs; and auditors or regulators who want a concrete example of how capability levels, evaluations and mitigation levels are meant to fit together. ## 📄 Source Details The document is titled *Frontier Safety Framework*, Version 1.0, and runs to seven pages in English. The acknowledgements name eight developers and list leadership including Allan Dafoe, Helen King, Tom Lue, Nicklas Lundblad, Demis Hassabis, Anca Dragan and Shane Legg, and credit METR for expertise on Responsible Capability Scaling. No publisher imprint, publication date or reference number is printed. Footnotes cite related work at gov.uk, METR, Anthropic and OpenAI, not this document. The full text of all seven pages was available; the extraction drops characters in some words.