Microsoft

Microsoft’s Azure Prompt Shield, deployed across its AI services including Azure OpenAI and other enterprise platforms, was revealed to harbor a critical security vulnerability through a deceptively simple emoji smuggling technique. Researchers from Mindgard and Lancaster University demonstrated that by embedding malicious instructions within Unicode emoji variation selectors, attackers can bypass the shield’s content inspection pipeline entirely. Because Azure Prompt Shield fails to normalize or parse these hidden characters in line with the underlying language model, it remains blind to the hidden payload while the model itself executes the commands. In controlled tests, this bypass achieved a perfect 100% success rate, enabling adversaries to unleash unauthorized code execution, data exfiltration attempts, and disallowed content generation. The implications are profound: enterprises relying on Azure’s guardrails may unknowingly expose sensitive intellectual property, customer data, and internal decision-making processes to hostile actors. This flaw not only undermines user trust in Microsoft’s AI safety infrastructure but also highlights an urgent need for more robust Unicode handling and unified guardrail-LM dataset alignment.

Source: https://cybersecuritynews.com/hackers-can-bypass-microsoft-nvidia-meta-ai-filters/

TPRM report: https://scoringcyber.rankiteo.com/company/microsoft

"id": "mic846050725",
"linkid": "microsoft",
"type": "Vulnerability",
"date": "5/2025",
"severity": "100",
"impact": "5",
"explanation": "Attack threatening the organization’s existence"

{'affected_entities': [{'industry': 'Technology',
                        'name': 'Microsoft',
                        'type': 'Technology Company'}],
 'attack_vector': 'Emoji Smuggling',
 'data_breach': {'data_exfiltration': True,
                 'type_of_data_compromised': ['Sensitive intellectual property',
                                              'Customer data',
                                              'Internal decision-making '
                                              'processes']},
 'description': 'Microsoft’s Azure Prompt Shield, deployed across its AI '
                'services including Azure OpenAI and other enterprise '
                'platforms, was revealed to harbor a critical security '
                'vulnerability through a deceptively simple emoji smuggling '
                'technique. Researchers from Mindgard and Lancaster University '
                'demonstrated that by embedding malicious instructions within '
                'Unicode emoji variation selectors, attackers can bypass the '
                'shield’s content inspection pipeline entirely. Because Azure '
                'Prompt Shield fails to normalize or parse these hidden '
                'characters in line with the underlying language model, it '
                'remains blind to the hidden payload while the model itself '
                'executes the commands. In controlled tests, this bypass '
                'achieved a perfect 100% success rate, enabling adversaries to '
                'unleash unauthorized code execution, data exfiltration '
                'attempts, and disallowed content generation. The implications '
                'are profound: enterprises relying on Azure’s guardrails may '
                'unknowingly expose sensitive intellectual property, customer '
                'data, and internal decision-making processes to hostile '
                'actors. This flaw not only undermines user trust in '
                'Microsoft’s AI safety infrastructure but also highlights an '
                'urgent need for more robust Unicode handling and unified '
                'guardrail-LM dataset alignment.',
 'impact': {'brand_reputation_impact': 'Undermines user trust in Microsoft’s '
                                       'AI safety infrastructure',
            'data_compromised': ['Sensitive intellectual property',
                                 'Customer data',
                                 'Internal decision-making processes'],
            'systems_affected': ['Azure OpenAI', 'Other enterprise platforms']},
 'lessons_learned': 'Need for more robust Unicode handling and unified '
                    'guardrail-LM dataset alignment',
 'motivation': ['Unauthorized code execution',
                'Data exfiltration',
                'Disallowed content generation'],
 'post_incident_analysis': {'root_causes': 'Failure to normalize or parse '
                                           'Unicode emoji variation selectors'},
 'title': 'Azure Prompt Shield Vulnerability',
 'type': 'Vulnerability Exploitation',
 'vulnerability_exploited': 'Unicode emoji variation selectors'}