Safety

Open Safety Framework

We are committed to the principle that the power of the most consequential technology of our time should be broadly shared, for all students, researchers, developers, companies, and institutions to study, build on, and deploy, rather than concentrated in the hands of a few organizations.

Introduction

The mission of Reflection AI, Inc. ("Reflection") is to make intelligence open and accessible to all. We build open models that let anyone control their intelligence and help shape the future of AI. We are convinced that openness accelerates progress: shared systems facilitate broader participation and produce ecosystems of research, innovation, and economic opportunities that no single lab can replicate alone. Open models let researchers inspect, stress-test, and strengthen the safeguards we build, contributing to the safety of the model and the entire ecosystem.

Reflection is committed to the principle that the power of the most consequential technology of our time should be broadly shared, for all students, researchers, developers, companies, and institutions to study, build on, and deploy, rather than concentrated in the hands of a few organizations. The most effective way to meaningfully control and govern an AI system or product is through access to open models, enabling organizations to operate AI under their own governance, security, and compliance frameworks.

Safety is at the core of Reflection’s mission. As our shared understanding of frontier risks continues to evolve, this framework represents Reflection’s current understanding and approach to ensuring our models, and the industry, measure dangerous capabilities and deploy open AI models responsibly.

Our Approach

Reflection’s approach is informed by our research and development of AI models, legal requirements, voluntary commitments such as the White House Accord on Super Intelligence, standards issued by national and international standards-setting organizations (including, but not limited to NIST AI RMF and ISO 42001), and industry norms and practices. Reflection anticipates that our approach will continue to evolve as legal requirements, standards, industry practices, and our research and development evolve as well.

To put these commitments into practice, Reflection has established a standing governance process with clearly assigned ownership, cross-functional representation across research, engineering, security, product, legal, and policy, and an auditable path for escalating the most consequential decisions to key leaders within the company and, where appropriate, Reflection’s Board of Directors.

Because openness is central to how Reflection builds, we welcome sustained engagement from the open-source community, government, researchers, and industry partners to test and continuously improve our models and safety systems. This collaboration will support our safety research, red teaming, model capability evaluations, open-weights safety systems, and evolving marginal risk landscape. It calls for scrutiny across the whole ecosystem to ensure the path to Superintelligence is guided by transparency, rigor, and shared accountability.

This Open Safety Framework (OSF) is focused on our approach to addressing risk for models that may exhibit ”catastrophic risk” as defined under relevant U.S. law, such as California’s Transparency in Frontier AI Act (TFAIA) and “general-purpose AI models with systemic risk” under the European Union’s AI Act. Where the OSF references “systemic risks” this includes that term, as defined under the EU AI Act, and “catastrophic risk” as defined under relevant U.S. law, such as the TFAIA. Reflection addresses other potential harms through separate policies, model specifications, and product safeguards.

This OSF is Reflection’s public commitment to identifying, evaluating, and mitigating systemic risks in our model. It sets out the specific risks to be addressed, the capability threshold that separates an ordinary release from one requiring additional mitigations, and the requirements that must be satisfied before Reflection releases any model.

Risks, Capabilities, and Mitigations

This Framework addresses eight recognized systemic risks and is constructed backwards from the harm using an outcome-first method. The first systemic risk, Concentration of Power, is evaluated using ecosystem-level indicators. The remaining seven are capability risks, each assessed for the specific catastrophic harm it could enable.

The Governing Logic

Release decisions balance two types of potential systemic risk:

Concentration risk: the potential loss of societal oversight that results from a powerful capability is held by only a few actors.
Diffusion risk: the potential misuse that may occur where a dangerous capability is released broadly.

Releasing a model openly reduces concentration harm but can raise diffusion harm; restricting or closing a model does the reverse. Of the two, concentration harm is the more durable: once structurally entrenched and self-reinforcing, it weakens the competition, institutions, and public oversight needed to challenge or reverse it. Diffusion harms, while potentially catastrophic or irreversible in individual cases, are more amenable at the system level to detection, adaptation, mitigation, and defensive response. Reflection’s default is therefore to release AI models openly but to maintain a backstop or baseline to address the most severe cases. This backstop applies where a capability's direct harm has the potential to be both catastrophic and irreversible. Reflection reaches that determination by weighing the capability's marginal risk, the offense-defense balance, and the potential for defensive acceleration.

Established Systemic Risks

In addition to the concentration systemic risk described above, the OSF takes into account four additional specified systemic risks.

Concentration of Power
Biological Weapons
Chemical Weapons
Cyber Offense
Loss of Control

Research Systemic Risks

The OSF also takes into account the following additional risks, which are separately categorized for the reasons described below.

Radiological Weapons
Nuclear Weapons
Harmful Manipulation

Thresholds and Response

This Framework establishes the conditions that govern Reflection’s response for each identified systemic risk. For a given systemic risk, Reflection assesses the relevant specific model and the proposals for its release and deployment to determine where that risk stands and the response it requires. This assessment is provisional and is reassessed following a material change.

Reflection operationalizes the safety margin through four factors assessed before release: (1) capability uncertainty, recognizing that evaluations may not fully characterize the model’s capabilities under all elicitation methods reasonably available at the time of assessment; (2) under-elicitation risk, recognizing that no evaluation exhaustively demonstrates model capability; (3) historical evaluation uncertainty, reflecting the observed limitations of comparable evaluation methodologies; and (4) mitigation-durability uncertainty, recognizing that credited mitigations should be demonstrated to remain effective under reasonable adversarial testing before release.

The Framework assesses Concentration of Power first (Table 1 below), using ecosystem conditions and release-mode indicators. The Established Systemic Risks that follow (Table 2 below) are each assessed against two risk tiers: Limited and Critical.

Concentration of Power is the systemic risk that frontier AI capability becomes concentrated in the hands of a small number of actors. Unlike the other systemic risks, it does not depend on misuse of the model. Reflection assesses whether the model’s capability and distribution increase or reduce that concentration and, where the risk is elevated, responds by promoting broader diffusion of frontier AI capability.

Table 1. Concentration of Power

Table 1
Element	Approach
Ecosystem	In a healthy ecosystem, frontier capability is broadly distributed across many open and closed developers, and no actor, including Reflection, holds or is nearing a decisive advantage. The condition of concern is concentration of frontier AI capability in the hands of a few companies, states, networks, or lab insiders, creating a durable position that democratic, legal, or market mechanisms can’t effectively reverse.
Indicators of concentration of power	Lack of distribution of capability across open and closed developers; limited breadth of effective access across the research and developer community; proximity of any single actor to a decisive advantage.
Response	Maintain open release as the default and monitor the ecosystem indicators. Where the indicators point toward a strong concentration of power, promote diffusion more strongly: release openly, and where Reflection holds the lead, release rather than hoard, even at competitive cost. Add measures to ensure broad effective access, publishing research and tooling and supporting downstream capacity, so that diffusion is real, not merely nominal. Withholding is not a response to concentration, since for Reflection that would mean becoming the concentrator.

Established System Risks are each assigned a direct-harm tier on a marginal basis, the additional risk beyond what comparable models already allow: either Limited (no meaningful uplift) or Critical (a decisive contribution to a catastrophic outcome).

Table 2. Established Systemic Risks

Biological Weapons
Tier	Capability	Mitigations
Limited	Limited expansion of misuse opportunities or the threat landscape relative to a reference model or public baseline, including comparable publicly available open-weight models or closed models without robust safeguards. The scale, sophistication, and accessibility of harmful activity remain substantially unchanged.	Pre-release biological capability evaluations, biosecurity SME red-teaming, and external review prior to release, calibrated to target misuse without suppressing legitimate life-science, diagnostic, and scientific capability.
Critical	Decisive contribution toward a Biological Outcome (a biological weapon capable of causing mass casualties deployed with weapons effect by a novice threat actor with the model as the material contributor). The contribution is decisive where the deployment would not have been possible with a released open-model or closed frontier model without robust safeguards.	Recommended deployment mitigations in the form of classifiers and intervention recommendations (system instruction intervention, prompt block, model routing, account summarization and/or monitoring). Release is paused until remediation demonstrates the recommended mitigations, offense-defense balance, defensive acceleration, threat environment, marginal risk, and model capabilities satisfy the Limited release criteria, supported by an affirmative safety case; otherwise, the model is not released.

Chemical Weapons
Tier	Capability	Mitigations
Limited	Limited expansion of (i) the misuse opportunity or (ii) the threat landscape for chemical agents. Expansion is marginal relative to a reference model or public baseline, including comparable publicly available open-weight models, or closed frontier models without robust safeguards where appropriate. The scale, sophistication, and accessibility of harmful activity are substantially unchanged.	Pre-release chemical capability evaluations, chemical-weapons SME red-teaming, and external review prior to release.
Critical	Decisive contribution toward a Chemical Outcome: a chemical weapon capable of causing mass casualties is deployed with weapons effect by a novice threat actor with the model as the material contributor, which would not have been possible with a released open-model or closed frontier model without robust safeguards.	Recommended deployment mitigations in the form of classifiers and intervention recommendations (system instruction intervention, prompt block, model routing, account summarization and/or monitoring). Release is paused until remediation demonstrates the recommended mitigations, offense-defense balance, defensive acceleration, threat environment, marginal risk, and model capabilities satisfy the Limited release criteria, supported by an affirmative safety case; otherwise, the model is not released.

Cyber Offense
Tier	Capability	Mitigations
Limited	No material contribution toward a Cyber Outcome: Limited or no material shift in the cyber-offense and/or cyber-defense balance attributable to the model. Limited expansion of the threat landscape relative to comparable publicly available open-weight models or closed frontier models without robust safeguards where appropriate.	Cyber refusal and alignment robustness training; red teaming; standard evaluations of autonomous vulnerability discovery and exploit generation; internal security controls; bug-bounty program.
Critical	Decisive contribution toward a Cyber Outcome: A. A severe cyber-attack on critical infrastructure resulting in casualties, catastrophic economic loss, or significant measurable disruption of said critical services, which would not have been possible with a released open-model or closed frontier model without robust safeguards. OR B. A significant advantage for cyber offensive capabilities, eroding defensive equilibrium, including dual-use categories (reverse engineering, penetration testing, etc.), which is not possible with a released open-model or closed frontier model without robust safeguards.	Recommended deployment mitigations in the form of classifiers and intervention recommendations (system instruction intervention, prompt block, model routing, account summarization and/or monitoring). Where the model can still conduct autonomous discovery-to-exploitation against critical-infrastructure systems after remediation resulting in stated effects, it is not released.

Loss of Control
Tier	Capability	Mitigations
Limited	No material erosion of evaluability, monitorability, or controllability attributable to the model. Within the elicitation methods and reference scaffolds tested, the model does not exhibit the properties that produce the Control Outcome: it complies with shutdown, interruption, tool restriction, and correction, including where compliance defeats an assigned task; does not attempt persistence beyond intended termination, self-replication, weight self-exfiltration, or egress from its authorized environment; and does not act to evade or deceive monitoring or evaluation. Comparison to a reference open-weight model informs this assessment but does not by itself establish the tier.	Mitigations are applied in proportion to the assessed risk and the release modality, and are drawn from the following: pre-release loss-of-control evaluations.
Critical	The model exhibits properties or attempts conduct that defeat evaluability, monitorability, or controllability: it persists, propagates, self-modifies, acquires resources, or resists correction beyond the operator's control, such that operators can no longer dependably assess its behavior, observe it in use, steer it, keep it within bounds, or bring it to a stop. Established by resistance to or subversion of shutdown or interruption; attempted or achieved egress, self-replication, weight self-exfiltration, or persistence beyond intended termination; action to evade, disable, or deceive monitoring, evaluation, or human oversight, including evaluation sabotage and capability concealment; or acquisition of resources, credentials, or permissions beyond authorized scope in service of the foregoing.	Recommended deployment mitigations, including monitorability classifiers over reasoning traces and tool calls, introspection probes for deception and evaluation awareness, and a reference agent harness with default egress restriction, tool allowlisting, and an out-of-band stop, are published and shipped enabled by default. They receive no credit toward this determination because a party holding the weights may remove them, and they address the mechanisms the model has been shown to defeat. The model is not released. Internal use of the checkpoint is suspended, and the matter is escalated to the Committee. Release may proceed only if remediation addresses the propensity in the weights and re-evaluation satisfies the Limited criteria under the same conditions, supported by an affirmative safety case. Otherwise, the model is not released in any modality.

Affirmative safety case. The safety case demonstrates that the residual systemic risk of the model remains acceptable after mitigations have been applied and documents: (1) the supporting evaluation evidence, credited mitigations, and applicable safety margin; (2) the assumptions and reasonably foreseeable conditions under which that conclusion would no longer hold; and (3) the rationale supporting the release decision.

Release gate. For each release candidate, the Model Report records the basis for proceeding, setting forth the following: (1) the systemic risk tier reached for each identified risk against the defined capability tiers, with the safety margin applied; (2) the systemic risk estimates, both with mitigations in place and with the model fully elicited, and the marginal risk the open-weight release adds relative to a similar model already available; (3) the safety and security mitigations implemented and their limitations; (4) whether and how input from independent external evaluators informed the decision, including their reports or a reasonable justification where no such external evaluators were involved; and (5) a reasonable justification that the systemic risks are acceptable, including the safety margins, and detailing the conditions under which that justification would no longer hold. The Committee (defined in the Governance section) proceeds only where the systemic risks are acceptable; otherwise, it does not place the model on the market, restricts or blocks it.

Research Systemic Risks

The Research Systemic Risks list comprises systemic risks Reflection does not yet assess against our standard release tiers. For these risks, bespoke evaluations may be required because the factors that drive harm extend beyond model capability alone. In addition, mitigations may not always be the best fit, and because some of these harms are diffuse across the ecosystem rather than tied to a single provider, both risk and response remain open research problems. We take these risks with the utmost seriousness and continue to study their marginal risk, capabilities, and threat models. Their placement as research systemic risks is provisional: we may add risks as capabilities advance and new threats emerge, and a risk might move to Established status once Reflection determines that tier-based assessment is warranted.

Table 3. Research

Table 1
Research Category	Description
Radiological Weapons	A radiological weapon capable of causing widespread contamination and disruption, deployed with weapons effect by a novice threat actor with the model as the material contributor, which would not have been possible with a released open-weight or closed frontier model without robust safeguards.
Nuclear Weapons	A nuclear weapon capable of causing mass casualties, deployed with weapons effect by a novice threat actor with the model as the material contributor, which would not have been possible with a released open-weight or closed frontier model without robust safeguards.
Harmful Manipulation	Significant erosion of informed consent or rational deliberation in consequential individual or collective decision-making, with the model as the material contributor, which would not have been possible with a released open-weight or closed frontier model without robust safeguards.

Risk Assessment and Threat Modeling

Reflection conducts model capability evaluations, release safety evaluations, recommended safety stack mitigation robustness evaluations (at Critical), third-party assurance evaluations, and evidence-based threat modeling to model each risk to inform said evaluations and launch criteria.

Internal use: The same systemic-risk assessment applies to internal use of frontier models, including research, operational systems, and internal products. Internal deployment is assessed against the same capability thresholds and safeguards as pre-release deployment. Behavior indicating degradation of evaluability, monitorability, or controllability is assessed under Loss of Control and triggers the same escalation and remediation process as a model crossing a pre-release threshold.

Capability forecasting: For each Established systemic risk, Reflection estimates when one of its models is expected to exceed the highest tier previously reached for that risk. Forecasts are expressed as time ranges or probability distributions together with their principal assumptions and uncertainties. They inform evaluation planning, mitigation development, governance prioritization, and resource allocation, and are reviewed whenever material new information becomes available.

Frontier-risk behavior standard. Reflection defines a model-level model behavior spec against which safeguards and residual risk are evaluated. This standard is not a general content policy, but rather describes the minimum intended behavior in frontier-risk contexts.

In frontier-risk settings, a Reflection model is expected to:

Decline a request, channel it toward a safe alternative, or limit itself to defense-oriented assistance whenever the request would materially advance any specified-risk scenario family.
Withhold instructions, tactics, or tool-use patterns that would help circumvent safety controls, evaluation, monitoring, or human oversight.
Not engage in deception, coercion, or the fostering of dependency, and not materially sway a user's beliefs or decisions in high-stakes situations.
Stay amenable to operator control in agentic evaluations, respecting shutdown, tool restrictions, monitoring, and correction.

Model Reports (referred to as “transparency reports” under relevant U.S. law, such as TFAIA) describe how we evaluate these behaviors under the mitigated model, representative downstream configurations, and Reflection’s official deployment stack, where appropriate. For public open-weight releases, only behaviors that remain durable in the released model receive mitigation credit.

Mitigations: Reflection applies durable mitigations proportionate to the identified systemic risks and intended release modality. Their effectiveness is evaluated as part of the systemic-risk assessment, and mitigation credit is determined according to the release-mode principles below.

Representative mitigations include training-data filtering, behavioral fine-tuning, capability ablation, machine unlearning, retraining, and staged pre-release access where appropriate. They also include the recommended deployment mitigations Reflection publishes for deployers to adopt: prompted classifiers, introspection probes, and supervised fine-tuned classifiers, with suggested interventions (system-instruction intervention, prompt block, model routing, monitoring, account enforcement, and/or hierarchical summarization). Documentation, safety kits, and acceptable-use guidance will support deployers.

Security: Reflection maintains appropriate security controls over internal infrastructure, training pipelines, and unreleased model variants until model parameters are publicly released or securely deleted. Equivalent controls may be substituted where they achieve the same security objective and are documented in the Model Report.

Governance

Framework and Accountability: Reflection maintains a governance framework that assigns clear and documented responsibility for managing the systemic risks arising from its AI models. The framework establishes accountability across four complementary functions: oversight, risk ownership, support and monitoring, and assurance. Together these functions ensure that systemic risks are identified, assessed, mitigated, and monitored throughout the model lifecycle, and that responsibility for each activity rests with an identified role or body.

Governance bodies. The structure is multilayered so that decisions are taken at the appropriate level. Operational decisions are made as close as possible to the relevant model, capability, or deployment by the leaders accountable for that work. The Governance Committee, chaired by the CEO, is Reflection’s executive decision-making body for Critical systemic-risk matters and provides strategic oversight of this Framework. The Governance Council is the cross-functional management body responsible for the Framework’s day-to-day operation, coordinating systemic risk assessments, mitigation activities, and routine governance decisions. Council Working Groups support the Governance Council by conducting technical assessments and reviewing specific models, risks, or releases. This structure provides a clear and auditable path for escalation while enabling leadership to set risk calibration, strategic direction, and executive oversight.

Where appropriate, the Governance Committee may provide information, assurance, and recommendations to the Board of Directors regarding the effectiveness of the AI governance framework, material systemic risks, significant incidents, significant changes to the OSF, and the adequacy of resources allocated to AI safety, security, and compliance activities.

Risk Tiering and Escalation: Within this framework, each horizontal function applies the tiering to its own area using the methods and evidence best suited to it, such as model evaluations, security testing, legal and regulatory analysis, and/or stakeholder and reputational assessment, while the tier definitions, the residual-risk basis, and the escalation rules remain shared across all functions.

Assignment and escalation. Each responsible team assigns a tier. The Council Working Groups rate and decide matters assessed as Limited. Critical matters are escalated to the Committee for review and decision, with the Council preparing the matter and recommendation.
Cross-functional differences. Where functions assess a matter at different tiers, the highest tier assigned governs until reassessment resolves it.
Uncertainty. Where evidence is incomplete, contested, or the matter sits on a boundary, it is assigned the higher tier until assessment resolves it downward.

Risk Ownership: The Safety Lead holds overall responsibility for managing the systemic risks arising from Reflection’s AI models. This responsibility spans model development and deployment activities, the implementation of mitigation measures, and the management of responses to serious incidents related to model use and development. To ensure decisions are made by those closest to the work, the Safety Lead may assign operational responsibilities to the research, engineering, security and product leaders accountable for particular models, capabilities, or deployments, and may delegate working-level decisions to the Governance Council. Even in areas where operational responsibility has been assigned, the Safety Lead will maintain ownership and accountability for the management of systemic risk.

Support and Monitoring: The AI Governance Lead supports and monitors the framework. The role operates independently from research and product delivery activities, which enables it to provide objective challenges. Its responsibilities include supporting the organization of governance processes, conducting systemic risk assessments, monitoring the implementation and adequacy of mitigation measures, and identifying gaps or weaknesses in the framework. The AI Governance Lead coordinates the Governance Council and reports to the Governance Committee, providing a direct line between working-level monitoring and executive oversight.

Assurance: Where appropriate, Reflection will also seek independent external assurance from qualified experts, evaluators, auditors, legal advisors, and/or other third parties, each of whom is independent of the framework's day-to-day operation.

Lifecycle

Safety is a lifecycle commitment, not a single checkpoint. Reflection assesses systemic risk before a model is released and stays engaged with it across the model’s life.

Material capability assessment. Where several candidate release checkpoints are likely to share a materially equivalent risk profile, Reflection may use a material capability assessment, comparing a candidate against the most recent fully assessed reference checkpoint on model capabilities, to avoid running a full systemic-risk assessment on each. Where the candidate shows no meaningful capability uplift, sits clear of a Critical threshold, and introduces no new risk-relevant affordance (such as a new modality or tool access), it may be covered by the reference checkpoint's safety case; otherwise a full assessment is conducted.

Post-release reassessment. Reflection conducts post-release monitoring appropriate to its AI model deployments, designed to assess whether systemic risk exceeds acceptable levels or updates are needed to Model Reports. For public open-weight models, reassessment concerns the original systemic-risk assessment and affirmative safety case rather than downstream deployments or derivative models of similar capability.

US voluntary government-testing buffer. The lifecycle calendar reserves up to 30 days before release to other trusted partners for (voluntary) U.S. government access if a model may be designated a “covered frontier model” under the June 2, 2026 Executive Order, “Promoting Advanced Artificial Intelligence Innovation and Security” (i.e., where the classified cyber-capability benchmarking process led by Treasury, NSA and CISA, with the NSA Director designating, would apply). This is a planning buffer and a collaboration path – not a mandatory licensing, preclearance, or permitting requirement.

Critical safety incident identification and response. Reflection maintains a process for identifying and responding to “critical safety incidents” under relevant U.S. law, such as TFAIA, or “serious AI incidents” under the EU AI Act (collectively “critical safety incidents”). Potential critical safety incidents are immediately escalated to the AI Governance Committee. The Committee determines the appropriate technical, governance, and regulatory response, including remediation, intervention, and external reporting where required by law. All critical safety incidents and significant near misses are documented and incorporated into subsequent systemic-risk assessments, Model Reports, and Framework updates where appropriate.

Reflection also investigates the following categories of reportable safety incident under the same escalation process:

Autonomous model behavior outside intended user direction;
Inadvertent release, theft, misappropriation, or unauthorized access to model weights;
Critical failure of technical or administrative controls protecting frontier models.

Where required, Reflection provides an independent third-party audit pathway consistent with applicable legal requirements.

Reflection reports reportable incidents in accordance with the applicable legal deadline. Documentation relating to covered models and reportable incidents is retained for the periods required under the applicable legal regime.

A Model Technical Report is prepared before each covered model is placed on the market and updated whenever required by this Framework or applicable law. It documents the model’s capabilities, intended uses, limitations, systemic-risk assessment, evaluation evidence, mitigations, affirmative safety case, and release decision. The documentation will be retained for the periods required under the applicable legal regime.

Substantial modification criteria. A frontier model undergoes a substantial modification where changes materially affect the capabilities, propensities, affordances, deployment context, threat landscape, the effectiveness of mitigations on which the systemic-risk assessment depends, or any other aspect of the systemic-risk profile documented in the most recent Model Report. A reassessment that changes the highest systemic-risk tier reached by any Reflection model always constitutes a substantial modification. Changes that do not materially affect the systemic-risk assessment do not require a new public disclosure.

Framework review and update. Once a covered model has been placed on the market, Reflection reviews this Framework at least annually, or sooner where credible indications suggest that it is no longer fit for purpose or is not being applied as intended. Material modifications approved by the AI Governance Committee are published within 30 days of approval, together with a justification describing the nature of and reasons for the change, a version number, the date of the modification, and a changelog summarizing the updates. Reflection will provide authorities with an updated version of the Framework where required by local law.

Conclusion

This Framework establishes Reflection’s approach to frontier and open model safety. It defines the systemic risks in scope, sets measurable capability thresholds, assigns release consequences to each tier, and requires an affirmative safety case before a model may be placed on the market.

The Framework is intended to be operational, not merely descriptive. It is implemented through model evaluations, recommended mitigations, red-team studies, threat modeling, external review, governance escalation, Model Reports, incident response, and post-release reassessment. Where evidence is incomplete, contested, or close to a threshold, the higher-risk classification applies until reassessment demonstrates otherwise.

Reflection will continue to update this Framework as model capabilities, evaluation methods, threat landscapes, legal requirements, and industry practice evolve. The objective is to enable the responsible release of increasingly capable frontier and open models by grounding release decisions in rigorous evaluation, durable mitigations, and transparent governance.

Try Beam today