A yr in the past, we launched the world to Cisco AI Protection, the trade’s first actually complete enterprise AI safety answer.
Within the yr since, AI expertise has developed at an unbelievable tempo, and the AI safety panorama has seen seismic shifts in parallel. Groups had been as soon as involved that their chatbots would possibly produce dangerous or delicate outcomes; immediately, they have to account for these considerations and likewise cope with the integrity of third-party AI parts, AI functions with entry to troves of delicate information, and compromised brokers engaged on an adversary’s behalf.
Our mission has all the time been to maintain our clients protected towards a quickly altering AI danger panorama. In the present day, we’re elevating the bar but once more with the most vital leap ahead for Cisco AI Protection since its debut to guard AI functions and brokers throughout three key areas: end-to-end AI provide chain safety, superior algorithmic crimson teaming, and real-time agentic guardrails.
Let’s take a more in-depth take a look at among the newest capabilities we’re delivering with this subsequent leap ahead.
Deliver end-to-end safety to the AI provide chain
The supply of third-party and open-source belongings has made AI growth extra accessible than ever. Builders haven’t any scarcity of choices; there are over 2 million fashions and 500,000 datasets on the Hugging Face platform alone.
Guaranteeing the integrity and safety of those externally sourced parts is crucial when utilizing them for enterprise AI functions. Fashions, libraries, datasets—and with the introduction of brokers, MCP servers, assets, and instruments, too—can all comprise vulnerabilities that compromise a broader AI system if unchecked.
Cisco AI Protection brings seamless provide chain scanning to the AI growth workflow, cataloging belongings and surfacing potential dangers earlier than they will undermine the inspiration of your AI functions.
AI Invoice of Supplies (BOM) connects to repositories to create a consolidated stock of AI belongings and decide their provenance. MCP Catalog extends this discovery to MCP servers throughout private and non-private registries throughout the group. These capabilities convey transparency and centralized governance for fashions, datasets, instruments, third-party dependencies, and different crucial AI assets.
To confirm that these belongings are secure to make use of, AI Protection scans mannequin recordsdata, MCP servers, and full repositories to uncover vulnerabilities, malicious insertions, and different latent dangers. Mannequin backdoors, executable code, and compromised instruments can flip AI functions and brokers from useful assistants into an adversary’s strongest instruments towards your small business. With AI Protection, provide chain scanning integrates seamlessly into the AI growth workflow.
Expertise the subsequent evolution in algorithmic AI crimson teaming
Whether or not you’re working with an open-source mannequin or growing your individual, it’s essential to understand how an AI software will carry out in actual world situations earlier than deploying. In any case, brokers are succesful not solely of manufacturing dangerous content material or revealing delicate information but additionally taking probably damaging actions when prompted by a nasty actor.
When Cisco AI Protection launched final yr, our proprietary algorithmic crimson teaming expertise set the usual for mannequin safety assessments. In-depth testing throughout a whole lot of security and safety subcategories would take a human analyst weeks to carry out; with AI Validation, we delivered this in minutes.
In the present day, we’re introducing the subsequent evolution in algorithmic AI crimson teaming. Our utterly redesigned AI Validation permits single and adaptive multi-turn testing for fashions and brokers with even broader multi-lingual assist.
This new interface isn’t simply simple on the eyes. It offers clear insights and rapid safety suggestions based mostly on findings from AI Validation assessments. It additionally weaves outstanding AI safety frameworks and requirements all through testing, together with these from NIST, MITRE, OWASP, and our all-new Cisco AI Safety and Security Framework as properly. This taxonomy offers a clearer understanding of adversary targets and total AI danger publicity at an enterprise-wide degree.
Defend your AI brokers towards threats in actual time
Agentic AI unlocks new potentialities for companies, however its entry and autonomy have to be balanced with strong safety measures. In any case, an AI chatbot could be manipulated into saying one thing dangerous; an AI agent could be manipulated into doing one thing dangerous.
Cisco AI Protection debuted with runtime guardrails that monitored interactions between a consumer and AI software to mitigate threats like immediate injection, delicate information leakage, and poisonous content material. Protection has constantly expanded over the past yr as our staff transforms risk intelligence alerts into in-product protections.
Brokers aren’t simply interfacing with customers, nevertheless; they’re making calls to instruments and extra assets to execute their given duties. That’s why AI Protection now inspects and protects MCP visitors in actual time, bringing complete runtime safety to advanced interactions between customers, brokers, and instruments.
Furthermore, we’re increasing the arsenal of runtime protections out there in Cisco AI Protection with purpose-built agentic guardrails. These are designed to handle a wholly new class of rising threats focusing on AI brokers. For instance, our Instrument Exploitation guardrail prevents adversaries from hijacking related instruments to steal delicate information or execute another dangerous end result.
Get began with Cisco AI Protection
We first launched AI Protection to instantly tackle fears about AI safety and allow daring, fearless innovation. A yr later, as brokers are reshaping what we consider is feasible, our mission stays unchanged.
What we’ve shared on this weblog are simply among the methods we’re driving safety on this agentic period. There’s much more to get into: availability throughout 4 international areas; assist for cloud, hybrid, and on-premises deployments with Cisco AI PODs and Cisco Safe AI Manufacturing unit with NVIDIA; and native integrations with Splunk, ServiceNow, and CI/CD techniques, simply to call a couple of examples.
For those who’re becoming a member of us at Cisco Dwell in Amsterdam, cease by our sales space and see what’s new with AI Protection in individual. As all the time, you’ll be able to see the newest by scheduling time with an professional from our staff too.
Most of the merchandise and options described herein stay in various levels of growth. The supply timeline of those merchandise and options is topic to alter on the sole discretion of Cisco.
