Why this ( agentic AI risks ) matters now
The AI for Good Global Summit main subject – agentic AI risks, hosted in Geneva by the International Telecommunication Union (ITU), has become a leading venue where AI’s social, ethical, and technical stakes are debated. On 7 October 2025, a replay session titled Delegated decisions, amplified risks: Charting a secure future for agentic AI featured Meredith Whittaker (Signal Foundation President, privacy advocate) and Kenneth Cukier (Deputy Executive Editor at The Economist).
The discussion focused on “agentic AI” — systems that don’t just respond to prompts, but act independently on behalf of users. Unlike chatbots, these AI agents can schedule meetings, access banking apps, send emails, and manipulate data — often requiring root-level access to devices or platforms.
This conversation matters because it highlights a core trade-off of modern AI: the convenience of delegation versus the erosion of digital self-sovereignty. For policymakers, developers, and business leaders, the stakes are immediate.
What Meredith Whittaker and Kenneth Cukier actually said
Key quotes and highlights
Meredith Whittaker stressed that agentic AI blurs the line between apps and operating systems:
“We’re talking about agents that require root access to your digital life — your messages, your calendars, your accounts. That’s a profound shift in control.”
She argued that without strict safeguards, agentic AI could normalize pervasive surveillance and amplify corporate control over personal data.
Kenneth Cukier framed the issue in governance terms:
“We’ve delegated decisions to algorithms before, but this is different — we’re giving them the keys to the house. The risks scale with the access we permit.”
Together, they called for gatekeeper accountability, regulatory guardrails, and a cultural shift toward privacy-by-design in AI development.
The core agentic AI risks, explained
The replay outlined several overlapping risks:
- Privacy erosion — AI agents need sweeping data access to function. Once granted, this access is hard to monitor or revoke.
- Root access — Some agents demand permissions that rival or exceed OS-level privileges, creating a massive attack surface.
- Surveillance risk — If AI agents funnel data to cloud providers or advertisers, user privacy effectively disappears.
- Loss of user control — Delegating to agents risks users no longer understanding (or controlling) decisions made on their behalf.
- Blurred OS boundaries — Apps versus operating systems are merging, with agents acting as “meta-apps” governing all others.
Real-world examples & analogies
To make this concrete, imagine:
- An AI agent with calendar + email access decides to reschedule a meeting, but shares sensitive details with a third-party system.
- An agent with banking app permissions initiates a “helpful” transfer that exposes account numbers.
- A voice-based agent with microphone access unintentionally streams private conversations into the cloud.
Each scenario shows how delegated autonomy increases convenience but also multiplies privacy exposures.
Industry response & technical safeguards
Developers and enterprise teams can mitigate these risks with technical design choices:
- Privacy-by-design — bake security principles into architecture, not as an afterthought.
- Permission models — require explicit, granular opt-in for each function, with clear user audit trails.
- Least privilege — grant only the minimal access needed for a task, never blanket permissions.
- Secure enclaves — process sensitive data locally where possible, reducing cloud exposure.
Whittaker emphasized that self-regulation alone won’t suffice. Guardrails must be enforced by policy and market pressure.
Policy & governance recommendations
The session suggested several regulatory priorities:
- Gatekeeper responsibilities — Platforms that provide agentic AI should be accountable for preventing misuse.
- Standards for permissioning — International standards bodies (like ITU) could define minimum safeguards.
- Auditability & transparency — Require independent audits of agentic AI access and data flows.
Quick policy wins over – agentic AI risks
- Mandate user-facing dashboards showing agent permissions.
- Require default “least privilege” settings for AI systems.
- Enforce clear off-ramps (easy ways to disable/revoke agent access).
What organizations should do next — a practical checklist
For product teams
- Map out all permissions your AI agent requests.
- Test user experience with revocation flows (can permissions be withdrawn easily?).
- Run privacy impact assessments before launch.
For security engineers
- Enforce least-privilege by default.
- Implement local-first processing for sensitive data.
- Create monitoring tools to log and audit agent actions.
For policy & legal teams
- Track evolving international standards from ITU, OECD, EU, and US regulators.
- Draft transparent disclosures for users.
- Prepare compliance playbooks before regulation arrives.
Agentic AI risks: balancing convenience with digital self-sovereignty
Agentic AI is not science fiction. It’s already arriving in enterprise productivity suites, consumer apps, and operating systems. The AI for Good replay makes clear: we must act now to shape these systems with user agency, privacy, and accountability at their core.
The choice is stark — either delegate responsibly with oversight, or drift toward a future where our digital brains belong to someone else’s servers.
Let’s connect!
Source: AI for good