Keak — agentic threat model

Not certain from the listing — The specific foundation models used by Keak for variant generation are not disclosed. The primary risk is model reprogramming or prompt injection that could force the model to generate malicious payloads or phishing content disguised as optimized website variants.

Keak scans live website pages and ingests traffic/conversion data. If an attacker can manipulate the target website's DOM or inject malicious metadata, they could poison the agent's input data, leading to corrupted variant generation or skewed optimization decisions.

The agent framework orchestrates scanning, variant generation, and live deployment of A/B tests. Vulnerabilities here include insecure tool integration where the agent's write access to the client's website DOM is abused to inject unauthorized scripts or bypass content security policies.

Keak integrates directly with client site builders, frameworks, and Slack. Compromise of Keak's hosting infrastructure or API keys could allow lateral movement, enabling attackers to push malicious updates directly to all connected customer websites.

Keak relies on traffic analytics and real-time alerts to select winning variants. A lack of robust anomaly detection could allow attackers to game the evaluation loop by sending artificial traffic to promote malicious or sub-optimal variants.

Not certain from the listing — There is no mention of specific security certifications (e.g., SOC2), role-based access controls, or human-in-the-loop authorization gates before publishing live website modifications, which increases the risk of unauthorized automated changes.

Not certain from the listing — While Keak integrates with Slack and external APIs, there is no explicit multi-agent orchestration described. The ecosystem risk is primarily focused on third-party API abuse and unauthorized actions taken via Slack integrations.