EqualAI CEO: Your AI assistant is acting in your name. Who signed off?

A Manifold public record counts 59 AI-agent incidents since July 2025; 42 were disclosed in the past four months, the median time to discovery was 70 days, and only 11 were detected by a security control. The commentary highlights unauthorized actions involving email, reservations and bank accounts, and argues for specific permissions, assigned responsibility, continuous testing and incident reporting. It also notes California liability law and proposed federal anti-hacking legislation, while emphasizing that companies must establish safeguards to earn consumer trust.

Analysis

The investable issue is not whether one assistant made a mistake; it is whether agentic features can scale without transferring unpriced liability to platforms. The near-term effect is likely higher friction—permission prompts, narrower tool access and slower rollouts—which could defer engagement or monetization rather than materially alter META’s consolidated economics today. The more important 6–18 month risk is asymmetric: one consequential incident or adverse legal precedent could force costly controls and create an overhang on agent launches. Conversely, auditable permissions, reliable revocation and clear remediation may become adoption advantages, so governance is also a moat investment, not only compliance expense.

META has direct product exposure through Muse, but the cited vulnerability was patched; this does not establish a persistent security weakness across Meta’s products. JPM’s resistance to unauthorized agents protects account integrity, while a durable refusal to support properly authorized agents could eventually cede customer interaction to third-party interfaces. The commentary’s incident count is a useful signal, not a measure of incident rates: disclosure is selective and the observed sample may rise as reporting improves. California liability rules and any federal legislation are catalysts, but proposal is not enactment. The contrarian point: stricter controls may slow adoption in the next few quarters yet improve consumer trust and long-run usage. No strong standalone directional trade is supported absent evidence of product, revenue or legal impact.

Key Decisions for Investors

• META: Do not short solely on these reports. Track Muse/agent rollout pace, independent security findings, user permission and revocation controls, and any material legal or regulatory action; reassess if these produce a launch delay or a company-disclosed financial exposure.
• JPM: Treat agent blocking as prudent account-risk management, not a near-term earnings catalyst. Monitor whether JPM supports authenticated, user-authorized agents; sustained exclusion could become a longer-term customer-interface and payments-distribution risk.
• Watch the next 1–3 months for concrete catalysts: enacted liability or incident-reporting rules, additional independently verified agent failures, and evidence that platforms are restricting features. A rise in disclosed incidents alone is not proof that underlying risk is worsening.
• If considering a sector expression, favor a diversified cybersecurity/identity-governance watchlist over a single-name bet until procurement demand or regulation translates into measurable bookings. The thesis weakens if agent access controls become standardized at low cost and reported incidents decline without adoption slowdowns.