Cisco
Cisco (Cisco Systems, Inc.) is an American multinational technology corporation headquartered in San Jose, California, engaged in the research, development, and sale of networking communications technology. Its core business encompasses network infrastructure equipment such as routers and switches, as well as solutions in cybersecurity, cloud computing, collaborative communications, and software-defined networking. Cisco's products and services are widely deployed across enterprises, governments, and service providers, making it a significant participant in internet infrastructure.
Core Facts
Timeline (last 90 days)
Cisco argues that the "one bug, one CVE" model is heading toward collapse, because when vendors use LLMs to scan code and fix vulnerabilities in bulk, it is impossible to assign a CVE to every internally fixed vulnerability; moreover, there are scenarios where something is exploited without a CVE, CVEs lag behind exploitation, and one CVE corresponds to multiple patches.
Attack acceleration began long ago and predates the emergence of large language models, a trend corroborated by Cisco's own reports, Mandiant's M-Trends, and Verizon DBIR data.
Cisco Vice President of Product Management Anton Chuvakin has 25 years of security research experience and previously served as a Gartner Research Vice President and a member of the Google Cloud CISO Office.
Security vendors such as Cisco and Trend Micro have repeatedly warned about exposure risks under OpenClaw's default configuration.
Jhulu Panath, Director of Product Management for Cisco's OutShift, and Cindy Chiu, Senior Product Manager for Duo Security, explained the challenges agents pose to IAM systems and the approaches to addressing them in a technical sharing session.
Cisco's management of Tokens is phased: first give employees loose Token access so everyone becomes familiar with it, then shift to a Token rationing system like a per-head quota—however many Tokens a team gets, it must deliver corresponding output.
Patel mentioned that the next architectural question is how to increase Token output efficiency by 20% to 60% at the same cost.
Token consumption monitoring should be placed at the same monitoring level as system resilience and guardrail violations, and an Agent's economic behavior should be treated on par with its behavioral boundaries.
Research by the Cisco team verified that even when an AI Agent's permissions are completely correct, it may still choose tools unrelated to the task to perform operations.
The solution proposed by the Cisco team is: before each tool call, use a small model to read the task description and the tool description and output a binary judgment of relevant or off-topic.
19 more timeline events
All Facts (20)
Research by SmartBear indicates that after reviewing more than 400 lines of code, a developer's defect detection rate drops sharply.
65%Unverified浏览器内核中最常见的安全漏洞类型包括内存安全漏洞（缓冲区溢出、释放后使用UAF）、类型混淆漏洞和竞态条件漏洞
90%UnverifiedCisco的Webex Hologram也在探索沉浸式远程会议技术
80%UnverifiedCisco argues that the "one bug, one CVE" model is heading toward collapse, because when vendors use LLMs to scan code and fix vulnerabilities in bulk, it is impossible to assign a CVE to every internally fixed vulnerability; moreover, there are scenarios where something is exploited without a CVE, CVEs lag behind exploitation, and one CVE corresponds to multiple patches.
50%UnverifiedCisco Vice President of Product Management Anton Chuvakin has 25 years of security research experience and previously served as a Gartner Research Vice President and a member of the Google Cloud CISO Office.
50%UnverifiedAttack acceleration began long ago and predates the emergence of large language models, a trend corroborated by Cisco's own reports, Mandiant's M-Trends, and Verizon DBIR data.
50%UnverifiedSecurity vendors such as Cisco and Trend Micro have repeatedly warned about exposure risks under OpenClaw's default configuration.
50%UnverifiedToken consumption monitoring should be placed at the same monitoring level as system resilience and guardrail violations, and an Agent's economic behavior should be treated on par with its behavioral boundaries.
50%UnverifiedCisco's management of Tokens is phased: first give employees loose Token access so everyone becomes familiar with it, then shift to a Token rationing system like a per-head quota—however many Tokens a team gets, it must deliver corresponding output.
50%UnverifiedThis study tested only one model family, Gemma 3.4b, and results may differ with a different model family.
50%UnverifiedThe Cisco team deployed 352 tools across 12 MCP servers for testing, covering diverse tool types and scenarios.
50%UnverifiedThe base model used by the Cisco team was Gemma 3.4b, which after three stages of prompt optimization, SFT, and GRPO reinforcement learning reached 96.13% end-to-end accuracy on their test set.
50%UnverifiedThe test tasks in this study are described as synthetic data, not derived from real user scenarios.
50%UnverifiedThe binary judgment framework in this study may not cover the large gray areas of 'partially relevant' and 'contextually relevant' cases in reality.
50%UnverifiedResearch by the Cisco team verified that even when an AI Agent's permissions are completely correct, it may still choose tools unrelated to the task to perform operations.
50%UnverifiedThe solution proposed by the Cisco team is: before each tool call, use a small model to read the task description and the tool description and output a binary judgment of relevant or off-topic.
50%UnverifiedCisco MCP Scanner支持本地YARA、LLM裁判、Cisco AI Defense API、行为代码分析器和VirusTotal多种分析器，YARA模式无需注册且数据不离开本机。
50%UnverifiedCisco Catalyst Center (formerly DNA Center) is Cisco's network management and automation platform, providing REST APIs for operations such as device discovery, configuration deployment, and troubleshooting.
50%UnverifiedThere are no temporary mitigations for this Cisco vulnerability—only installing the patch is viable, and virtual appliances are recommended to be rebuilt from scratch.
50%UnverifiedThe vulnerability resides in the processing logic of Cisco AsyncOS for inbound emails; due to a lack of input sanitization, an attacker can execute code and gain root privileges by sending an email.
50%