Search
  • Home
  • Security teams can give AI more room to act, provided the limits are built into the system
  • Security teams can give AI more room to act, provided the limits are built into the system

    Sasha Slankamenac of Dariel

    Attackers are using AI to speed up reconnaissance, write more convincing phishing emails and coordinate parts of an attack. Defenders have access to much of the same tooling, and many security teams are now working out how far they can let it investigate and respond to incidents without a person signing off on each step.

    Sasha Slankamenac, Architect in the Office of the CTO and Practice Lead: AI at Dariel, points out that security systems have acted on their own for a long time. Firewalls block traffic automatically, and orchestration tools have been running predefined playbooks for years.

    “What generative AI adds is the ability to read context, pull evidence together from different sources and choose between possible actions,” he says. “So, before anything gets switched on, someone has to have decided what the system can do without asking first, and that needs people who understand both the business and the risk.”

    Where AI is helping security teams today Most of the value so far sits in the work between an alert firing and someone deciding what to do about it. A typical investigation involves pulling logs from several tools and checking indicators against threat intelligence, and much of an analyst’s time goes on that before any real assessment starts.

    “An analyst can spend most of an hour gathering information before making a judgement call,” says Slankamenac. “An agent can do a lot of that gathering, correlate the endpoint and identity logs, say, and give the analyst a summary with the evidence attached. It helps with triage as well, because obvious false positives can be closed off with the reason recorded, and related alerts get grouped into one incident.”

    Narrow containment steps, such as isolating a single laptop or blocking a known malicious domain, can be taken within seconds when the evidence is clear. Slankamenac says this is where speed counts most, since early detection achieves very little if the response still must wait for manual approval. What AI changes for analysts and their workload

    Experienced security people are hard to find and expensive to keep, and Slankamenac sees AI as a way of using their time more carefully. “A lot of the enrichment and write-up work, and honestly the fifth near-identical phishing report of the morning, can be handled by the system.

    That leaves analysts more time for the cases that need judgement, and for tuning the rules the AI follows.”

    He doesn’t see it as a route to smaller teams. “You still need experienced people deciding what the agent may do and checking its work, and handling whatever it escalates. Without them, nobody can really review what it’s doing.”

    Matching AI’s authority to the consequences of its actions Slankamenac’s working principle is to automate according to consequence. Actions with limited impact that can easily be undone, like gathering evidence or isolating one endpoint, can run with little or no approval. As actions become harder to reverse, reach further across the environment, or change who has access to what, the AI’s authority should narrow. Some of those actions need explicit approval, and some should be off-limits to the agent entirely.

    Requiring a human click on every step tends to bring back the delay that automation was meant to remove. “You can automate the whole investigation and still have the response sitting with one person waiting to approve it,” he says. “I’d rather people spent their judgement on setting the policy and thresholds, and on approving the decisions that carry real weight.”

    Deciding which actions fall into which category involves the wider business too. “Leadership needs to say which processes the organisation can’t afford to have interrupted, and which can tolerate an agent acting immediately, accepting that maybe one or two percent of those actions will turn out to be false positives that someone has to reverse by hand. The security specialists then configure the policies and controls that enforce that mandate.”

    How AI security tools can be manipulated through what they read Once an AI agent has a defensive role, attackers will try to target it. It reads emails, logs, tickets and web content, some of which may have been created or altered specifically to mislead it.

    “If an attacker can get text in front of the agent, they can try to slip instructions into it or push it towards a particular action,” Slankamenac says. “It needs the same protection you’d give any privileged component.” In practice, that means treating each input according to how far it can be trusted and keeping the agent’s permissions narrow, with its tool calls monitored so decisions can be audited afterwards.

    Putting the limits in identity, permissions and workflow

    Slankamenac says the model can’t be the final authority on what the model is allowed to do. “You can write ‘don’t disable a privileged account without approval’ into the prompt, and the agent will probably follow it. If it really matters, though, the agent shouldn’t have that permission available to it without an approval step built in.”

    If an agent should never delete infrastructure, the identity it runs under shouldn’t hold delete permissions, and where an action needs sign-off, the workflow should refuse to go ahead without it. The agent should only have access to the tools its role needs, with each call logged in enough detail to show afterwards what it did and why.

    For CISOs, this means AI governance has to show up in how the system is built. Before deploying a security agent, Slankamenac suggests checking whether the organisation can actually say which systems the agent can reach, and where it has to stop and escalate.

    “If those answers are vague, the authority hasn’t been properly defined yet, and it’s worth sorting that out before anything goes live,” he says.

    Portrait supplied by Dariel, reframed for publication.

    Facebook
    Twitter
    LinkedIn
    Pinterest