Three insights into agent security and cyber resilience

Agentic AI security tests where companies place their trust as software gains the ability to access sensitive data and act on behalf of users. The challenge ranges from controlling individual requests to controlling actions that can span multiple systems and business processes.

This expanding reach makes security an architectural and operational concern, with access, monitoring and recovery decisions increasingly interconnected. According to Krista Case of theCUBE Research, companies need to consider how these decisions impact their ability to maintain operations despite disruptions.

“With agent AI, we have these new agent-driven attacks that are making all the headlines,” she said. “When it comes to cyber resilience and business resilience, the result is that it has really become part of the security discussion. [Resilience] It’s important to build it in advance, knowing we can’t just build the fences high enough.”

Case and fellow analyst Dave Vellante explored these pressures during theCUBE’s coverage of Oracle’s virtual event, “AI Cyberattacks Escalate: How to Secure Your Data Now.” The discussions brought together Oracle executives and technology experts to examine the changing threat landscape and its impact on organizations. (*Disclosure below.)

Here is theCUBE’s full video discussion with Dave Vellante and Krista Case:

Here are three takeaways you may have missed from theCUBE’s coverage of Oracle’s “Escalating AI Cyberattacks: How to Secure Your Data Now” virtual event:

Insight #1: Data security must be enforced across applications and AI agents.

Agentic AI security requires a consistent authorization boundary as agents create new paths to corporate data. According to Juan Loaiza, executive vice president of Oracle Database Technologies at Oracle, Oracle Corp.’s approach is shifting. the enforcement of individual applications into the database, reducing the dependency on developers to implement protection measures separately.

“Humans are very clever at formulating things in a way that makes the AI ​​do things that you would have essentially thought that you would have prevented by asking the AI ​​to do,” he told theCUBE. “If we enforce at this data layer that each end user can only see their data, then the agent acting on that user’s behalf will only be able to see that user’s data, regardless of the type of prompt.”

Moving enforcement to the database also changes the way organizations manage security: protection must apply regardless of which application or agent requests access. “Oracle combines controls applied to database access with centralized security management, including cloud-based Oracle Data Safe and self-managed Oracle Database Security Central,” said Vipin Samar, senior vice president of database security at Oracle, who appeared with David Knox, vice president of database security product management at Oracle.

“AI can move from one database to another very easily and naturally,” Samar said. “We have to have [a] full 360 degree view [the] entire fleet, all your users, all your data. Only then can you have truly effective security.”

For partners building agent-based applications, this architecture separates the queries an agent generates from the data it is allowed to retrieve. Accenture Enkitec Group used Oracle Deep Data Security to enforce access rules within the database, allowing agents to query a common knowledge repository while maintaining different levels of authorization, said Roger Cornejo, principal director of technology innovation in Gen AI/AI at Accenture Enkitec Group.

“We have dozens, probably 40 or more agents using this application,” Cornejo told theCUBE. “We had to make sure it was also scalable. Right now we’ve tested it to scale to 300 concurrent queries and deliver most results in 15 seconds or less.”

Here is the full video from theCUBE with Vipin Samar and David Knox:

Insight #2: Agent AI security depends on prepared systems and responsible oversight.

To support agents, companies must modernize the systems they rely on. According to Holger Mueller, vice president and principal analyst at Constellation Research Inc., delaying adoption leaves companies exposed to AI-powered threats while technical debt increases.

“This year everyone is arguing about which is the better agent framework,” he said. “Next year will be the big renaissance of scaling backend systems.”

Readiness also includes knowing where agents are set up and what sensitive assets they can reach. According to Johnnie Konstantas (pictured), group vice president, Oracle Cloud Engineering at Oracle, this inventory helps companies establish controls for delegated work.

“Agent AI brings to the fore the idea that an agent and multiple subagents, acting as proxies for a person, can access data,” she told theCUBE. “So identity takes a new form in that the agent carries with them these access and access permissions [can] Maybe you even take action on a specific workflow. Make sure you have a good identity, as long as it is consistent and ubiquitous.”

These technical controls require organizational support. According to Brennan Baybeck, senior vice president and chief information security officer of Oracle Customer Success at Oracle, basic governance, security fundamentals and clearly understood customer and vendor responsibilities remain gaps in enterprise AI implementations.

“Humans have never been more important in controlling these robots as we call them,” he told theCUBE. “We used to say humans were the weakest link. But now I say robots are the most powerful and potentially dangerous link, and humans are now an essential link.”

Here is theCUBE’s complete video interview with Holger Müller:

Insight #3: Cyber ​​resilience requires trustworthy recovery and continuous improvement.

According to Ashish Ray, senior vice president of Oracle Database, product management at Oracle, agent AI security must take into account the possibility that both production systems and backups become targets. Separate management controls and network separation protect recovery resources, while Oracle’s Zero Data Loss Recovery Appliance verifies backup integrity and recoverability.

“Data that is secured but cannot be validated is Schrödinger’s data,” he said during the event. “The data could be backed up and you can’t rest on your laurels because the moment you try to restore, restore and use the data, you may find that the data cannot be used because there may have been some underlying validation checks and underlying corruption.”

Security programs must also consider that attackers discover and exploit vulnerabilities faster than people can test, approve, and deploy fixes. According to Zulfikar Ramzan, former chief technology officer and chief digital officer of RSA Security USA LLC, defenders’ knowledge of critical business assets helps prioritize protection and remediation.

“Continuously learn from every incident that occurs,” he told theCUBE. “Look at the near misses where you got away scot-free but maybe you shouldn’t have. Take that input, take that insight [and] Share it with your security program on a regular basis.”

Here is the complete video interview with Ashish Ray:

To see more of theCUBE’s coverage of Oracle’s “Escalating AI Cyberattacks: How to Secure Your Data Now” virtual event, check out our full video playlist here:

https://www.youtube.com/watch?v=videoseries

(*Disclosure: TheCUBE is a paid media partner for Oracle’s virtual event, “AI Cyberattacks Are Escalating: How to Secure Your Data Now.” Sponsors of theCUBE’s event coverage have no editorial control over theCUBE or SiliconANGLE content.)

Photo: SiliconANGLE