SHERIDAN, WY — June 10, 2026 — Safeguard today announced universal AI model support across every major frontier model family, making its Multi-Agent TAOR Deep Think AI Engine the first agentic security platform to offer customers a genuine choice of reasoning backbone for zero-day vulnerability discovery and automated remediation. Effective immediately, customers can run Safeguard's agentic pipeline on any combination of Anthropic Claude Fable 5, Claude Opus, Claude Sonnet, OpenAI GPT models, Google Gemini, Microsoft Phi, Meta Llama, and their own privately hosted models.
The announcement extends Safeguard's existing integration of Anthropic's Mythos-class models — including Claude Fable 5, released yesterday — and responds to enterprise demand for model flexibility in production security programs, where organizational AI policy, data residency requirements, cost structures, and security clearance levels each impose different model constraints.
"Every serious security team has a model strategy," said Hritik Kumar Sharma, Founder and CEO of Safeguard. "Some are Anthropic shops. Some run GPT-5.5 across their engineering stack. Some have a Gemini enterprise agreement. Some are building their own models on proprietary security data. Safeguard should work for all of them — and as of today, it does."
Supported Model Families
Safeguard's agentic zero-day discovery and remediation pipeline now supports the following model families natively:
Anthropic Claude Claude Fable 5, Claude Opus 4.8, Claude Sonnet 4.6, Claude Haiku, and Claude Mythos Preview for Project Glasswing partners. Safeguard has been an Anthropic Project Glasswing partner since April 2026 and ships same-day support for new Claude releases.
OpenAI GPT GPT-5.5 and the full GPT-4 family, including function-calling and structured output formats optimized for Safeguard's multi-agent orchestration layer. Enterprise customers with existing OpenAI agreements can route Safeguard workloads through their contracted capacity.
Google Gemini Gemini 3.1 Pro, Gemini 3.1 Flash, and Gemini Ultra via Google Cloud Vertex AI and Google AI Studio. Gemini's context window and multimodal capabilities integrate with Safeguard's visual supply chain analysis pipeline for infrastructure diagram and architecture review.
Microsoft Phi and Azure OpenAI Microsoft Phi-4 and Azure OpenAI deployments, including models running inside Azure Government and Azure for Operators environments. This enables Safeguard deployments in regulated public sector and defense-adjacent contexts that require Azure-bounded data handling.
Meta Llama Meta Llama 4 and Llama 3.3 families, including self-hosted, cloud-managed, and fine-tuned variants. Llama support is particularly relevant for organizations that require fully air-gapped deployments where no model inference can cross a network boundary.
Safeguard Native Models Safeguard's own Griffin Zero, Griffin L, Griffin M, Griffin S, and Griffin Lite — the security-only model lineup trained exclusively on cybersecurity data, available for sovereign, air-gapped, and on-premises deployments.
Private Model Hosting: Bring Your Own Model
For organizations that have developed their own AI models — whether fine-tuned on proprietary security corpora, trained on internal codebase history, or purpose-built for specific vulnerability classes — Safeguard now offers private model hosting within its platform infrastructure.
Under the private model hosting program, a customer's model is deployed into an isolated, policy-governed inference environment within Safeguard's platform. That model then becomes a first-class agent in Safeguard's multi-agent pipeline, participating in zero-day discovery, verification, and remediation workflows alongside Safeguard's native agents and any frontier models the customer has configured.
This means an organization that has fine-tuned a model on five years of their own CVE history, proprietary exploit data, or internal code patterns can use that model as their primary reasoning agent — with Safeguard's verification pipeline, supply chain context, and remediation automation wrapped around it. The proprietary model stays within the customer's data boundary; Safeguard handles the orchestration.
"Organizations building internal AI security capability have been forced to either abandon the models they have built or rebuild the entire orchestration stack themselves," said Hritik Kumar Sharma. "Private model hosting in Safeguard means your model is the agent — and the entire TAOR Deep Think pipeline becomes its reasoning infrastructure."
Model-Agnostic Agentic Architecture
All supported models — whether Anthropic, OpenAI, Google, Microsoft, Meta, Safeguard native, or privately hosted — operate as agents within the same Multi-Agent TAOR Deep Think AI Engine. The architecture does not treat model choice as a configuration option on top of a fixed pipeline. Each model operates with its own tool access grants, policy boundaries, and verification responsibilities within the multi-agent graph.
Customers can mix model families within a single deployment: for example, routing initial zero-day discovery through Claude Fable 5 for its FrontierCode Diamond-tier code analysis capability, passing candidate findings to a custom internal model for organization-specific context enrichment, and routing remediation generation through GPT-5.5 for teams whose developers are already calibrated to GPT output style. Each agent in the graph operates at the capability level of the model powering it. The system as a whole reflects the combined reasoning of the full agent team.
Verification, exploitability assessment, severity contextualization, supply chain blast radius mapping, and remediation pull request generation all run consistently across model choices. The output is a confirmed, contextualized, remediation-ready finding — regardless of which model family produced the initial detection.
Continuous Zero-Day Discovery Across All Model Families
Zero-day discovery in Safeguard runs continuously against every code change, on every model configuration. The platform does not require customers to choose between model capability and coverage cadence. A customer running GPT-5.5 as their primary agent gets the same continuous, per-commit discovery cadence as a customer on Claude Fable 5 or a customer running a private Llama fine-tune.
The model family determines the depth and character of the reasoning. The platform architecture determines the reliability, verification quality, and remediation completeness. Both are independently configurable.
Availability
Universal model support is available immediately for all Safeguard customers. Private model hosting is available under an enterprise agreement. Organizations interested in private model hosting are encouraged to contact press@safeguard.sh to discuss deployment requirements.
About Safeguard
Safeguard is the agentic software supply chain security platform that fuses 11 scanners, a security-only AI model lineup (Griffin · Eagle · Lion), universal frontier model support, and reachability-aware multi-agent reasoning to find what pattern scanners miss — from CVEs to candidate zero-days — and to ship the fix with cited reasoning. Customers can run the platform on Anthropic Claude, OpenAI GPT, Google Gemini, Microsoft Phi, Meta Llama, Safeguard's native models, or their own privately hosted models. Learn more at https://safeguard.sh.
Media Contact
press@safeguard.sh