Visual Exclusivity Attacks: Automatic Multimodal Red Teaming via Agentic Planning
2026-03-24T07:23:35Z•08319f0d0cee3bb2cfb64988535ff6f54b1afddde9c3e826489d90063d780310
access-control-agentsagentic-planningbounds-protectioncheckpoint-restorefederated-learning-securitygpu-reproducibilityhomomorphic-encryptionjailbreakingllm-agentsmemory-poisoningmemory-safetymobile-ad-detectionmulti-agent-systemsmultimodal-attackssemantic-rollbacksupply-chain-verification
What happened
This corpus (multiple arXiv submissions) highlights emerging security risks and defenses at the intersection of ML/agents, systems, and cyber security. Key themes: (1) Multimodal agentic attacks — "Visual Exclusivity (VE)" and MM-Plan show that agentic, multi-turn planning over images can jailbreak/abuse state-of-the-art multimodal LLMs (46.3% success vs Claude 4.5 Sonnet; 13.8% vs GPT-5), revealing a gap in visual reasoning defenses. (2) Agent/agent-system threats — memory poisoning for agentic and multi-agent systems is analyzed with proposed mitigations (cryptography, private knowledge-retr
Why it matters
A reviewed impact interpretation has not been published for this record.
Evidence and limitations
- Source ID
- arxiv_cs_cr
- Record identifier
- 08319f0d0cee3bb2cfb64988535ff6f54b1afddde9c3e826489d90063d780310
- Enrichment time
- 2026-03-24T07:23:35Z
- AI-assisted enrichment
- Yes
This record may overlap with other records. Its enrichment can be incomplete or wrong, and machine assistance was used. Validate consequential decisions against the linked source and your own environment.