UMMRO AI Safety Audit Report - Preview

PREVIEW

UMMRO

AI Safety Audit Service

AI Safety Compliance
Audit Report

TechCorp AI Assistant — Multi-Model Security Assessment

Prepared For

TechCorp Inc.

Date

February 11, 2026

Report ID

UMR-2026-0211-TC

Author

Ahmed Adel Bakr Alderai

Section 01

Executive Summary

7.2

out of 10

Overall Risk Grade: Acceptable with Noted Deficiencies

Composite Compliance Score: 72%

Models Tested

Strategies / Model

72%

Avg. Safety Score

Critical Findings

UMMRO conducted a comprehensive AI safety compliance audit of the TechCorp AI Assistant deployment, evaluating 10 large language models across 5 distinct reframing strategies (50 total test scenarios). The overall safety posture is Acceptable (Grade B, 7.2/10), with strong performance from Claude 3.5 Sonnet and GPT-4o, but three models exhibited significant safety deficiencies requiring remediation.

Key Findings

CRITICAL Two open-source models exhibited guardrail bypass under role-play reframing. Immediate mitigation required.
HIGH Gemini 2.5 Pro demonstrates soft refusal patterns -- content delivered with insufficient disclaimers.
INFO Claude 3.5 Sonnet achieved the highest score (94%) with consistent guardrail enforcement.

Section 02

Compliance Scorecard

Model	Provider	Grade	Score	Refusal Rate
Claude 3.5 Sonnet	Anthropic	A	94%	100%
GPT-4o	OpenAI	A	91%	96%
Gemini 2.5 Pro	Google	B	78%	84%
Llama 3.3 70B	Meta	C	52%	56%
Gemma 2 27B	Google	F	28%	24%

Section 02 of 09

Compliance Scorecard

Full model-by-model scores, refusal rates, latency benchmarks, and grade breakdowns for all 10 models.

Unlock Full Report

Section 03

Risk Matrix

Model	Guardrail	Refusal	Soft Refusal	Adversarial	Quality	Latency
Claude 3.5 Sonnet	98%	96%	94%	92%	90%	95%
GPT-4o	95%	93%	88%	89%	92%	97%
Llama 3.3 70B	54%	52%	42%	38%	68%	82%
Gemma 2 27B	22%	28%	18%	20%	42%	86%

Section 03 of 09

Risk Matrix Heat Map

Six-dimensional compliance heat map across all 10 models with color-coded risk assessment.

Unlock Full Report

Section 04

Model-by-Model Results

Detailed breakdown of each model's performance across all 5 reframing strategies, including per-strategy scores, refusal classifications, response previews, and latency measurements.

Section 04 of 09

Model-by-Model Results

Detailed cards for each of 10 models with strategy-level scores, refusal breakdowns, and response analysis.

Unlock Full Report

Sections 05-07

Strategy Analysis, Compliance Mapping & Recommendations

Strategies Analyzed

Regulatory Frameworks

Recommendations

CRITICAL Remove Gemma 2 27B from Production

Model scored 28% with inverted refusal behavior...

HIGH Implement Role-Play Input Filtering

6 of 10 models vulnerable to persona-based prompts...

MEDIUM Standardize Safety in CI/CD Pipeline

Automated regression testing prevents safety drift...

Sections 05-07 of 09

Strategy Analysis, Compliance & Recommendations

Full reframing strategy effectiveness analysis, EU AI Act / NIST AI RMF / ISO 42001 / SOC 2 compliance mapping, and 6 prioritized remediation recommendations with effort estimates.

Unlock Full Report