A brand new analysis led by LatticeFlow AI, in collaboration with SambaNova, gives the primary quantifiable proof that open-source GenAI fashions, when outfitted with correct threat guardrails, can meet or exceed the safety ranges of closed fashions, making them appropriate for implementation in a variety of use circumstances, together with highly-regulated industries equivalent to monetary companies.
Additionally Learn: AiThority Interview with Tim Morrs, CEO at SpeakUp
“At LatticeFlow AI, we offer the deepest technical controls to judge GenAI safety and efficiency,” mentioned Dr. Petar Tsankov, CEO and Co-Founding father of LatticeFlow AI.
The analysis assessed the highest 5 open fashions, measuring their safety earlier than and after making use of guardrails to dam malicious or manipulative inputs. The safety scores of the open fashions jumped from as little as 1.8% to 99.6%, whereas sustaining above 98% high quality of service, demonstrating that with the suitable controls, open fashions are viable for safe, enterprise-scale deployment.
Rethinking Open-Supply GenAI for Enterprise Adoption
Many corporations are actively exploring open-source GenAI to realize flexibility, scale back vendor lock-in, and speed up innovation. However regardless of rising curiosity, adoption has usually stalled. The rationale: an absence of clear, quantifiable insights into mannequin safety and threat.
The evaluations launched handle that hole, offering the technical proof wanted to make knowledgeable choices about whether or not and easy methods to deploy open-source fashions securely.
“Our clients — from main monetary establishments to authorities companies— are quickly embracing open-source fashions and accelerated inference to energy their subsequent technology of agentic functions,” mentioned Harry Ault, Chief Income Officer at SambaNova. “LatticeFlow AI’s analysis confirms that with the suitable safeguards, open-source fashions are enterprise-ready for regulated industries, offering transformative benefits in value effectivity, customization, and accountable AI governance.”
“At LatticeFlow AI, we offer the deepest technical controls to judge GenAI safety and efficiency,” mentioned Dr. Petar Tsankov, CEO and Co-Founding father of LatticeFlow AI. “These insights give AI, threat, and compliance leaders the readability they’ve been lacking, empowering them to maneuver ahead with open-source GenAI safely and confidently.”
Key Findings from the Analysis
LatticeFlow AI evaluated 5 broadly used open basis fashions:
- Qwen3-32B
- DeepSeek-V3-0324
- Llama-4-Maverick-17B-128E-Instruct
- DeepSeek-R1
- Llama-3.3-70B-Instruct
Every mannequin was examined in two configurations:
- Base mannequin, as sometimes used out-of-the-box
- Guardrailed mannequin, enhanced with a devoted enter filtering layer to dam adversarial prompts
The analysis centered on cybersecurity dangers, simulating enterprise-relevant assault eventualities (equivalent to immediate injection and manipulation) to measure every mannequin’s resilience and its influence on usability.
Key outcomes:
- DeepSeek R1: from 1.8% to 98.6%
- LLaMA-4 Maverick: from 33.5% to 99.4%
- LLaMA-3.3 70B Instruct: from 51.8% to 99.4%
- Qwen3-32B: safety rating elevated from 56.3% to 99.6%
- DeepSeek V3: from 61.3% to 99.4%
All fashions maintained over 98% high quality of service, confirming that safety good points didn’t compromise consumer expertise
Why This Issues for Monetary Establishments
As GenAI strikes from experimentation to deployment, enterprises face rising scrutiny from regulators, boards, and inside threat groups. Fashions should now be auditable, controllable, and provably safe.
This analysis gives clear, quantifiable proof that open-source fashions can meet enterprise-grade safety expectations with the suitable threat mitigation methods.
Additionally Learn: Cognitive Product Design: Empowering Non-Technical Customers By Pure Language Interplay With AI-Native PLM
[To share your insights with us as part of editorial or sponsored content, please write to psen@itechseries.com]
