TikTok CEO: Users Give AI Agents Dangerous Permissions They Do Not Understand
TikTok chief Shou Zi Chew warns that users blindly hand AI agents credit cards, creating massive security risks for platforms.
Business · Source: CNA
What happened
TikTok CEO Shou Zi Chew issued a stark warning at the DBS Asia Future Forward conference in Singapore. He said users are blindly granting agentic AI models permissions they do not comprehend. Chew used a specific example to illustrate the danger. If you give an AI agent access to your credit card to buy something, an overly eager agent might buy far more than you actually need. He stressed that cybersecurity risks tied to these agents are very real and require immediate user education.
These agents are fundamentally changing platform security. Unlike passive chatbots that just answer questions, agents execute tasks autonomously. Chew noted this dramatically lowers the cost for bad actors to generate spam and security threats. His solution is economic rather than purely technical. Platforms cannot reduce spam to zero, but they must build defenses that make abuse too expensive for attackers to sustain. Businesses need the right tools to defend against this fast-falling cost of attacks.
DBS CEO Tan Su Shan echoed the need for strict guardrails so agents do not run amok. She warned companies against applying AI lipstick by just bolting models onto old workflows. Meanwhile, TikTok is actively deploying these systems internally. Chew revealed TikTok now deploys thousands of AI agents overnight to audit safety feeds, replacing manual human testing. They also use vision language models to police live streams, reading details like speedometers that human moderators miss.
Key facts
- 10 to 12 — Major AI model releases happening per month according to DBS CEO Tan Su Shan
- 1,000 — Number of AI agents TikTok deploys overnight for youth safety checks
- $1 to $10 — Cost to produce a video ad using ByteDance's Seedance model
- 30 percent — Amount of an engineer's time actually spent coding
- 2.21 billion — TikTok's worldwide user base as of April
Why it matters
The shift from passive chatbots to active agents breaks existing security models. When AI can take actions on behalf of a user, the blast radius of a mistake or a hack multiplies exponentially. Builders must design permission systems that humans can actually understand at a glance. If your product requires users to hand over their wallet or core data access to an agent, you own the liability when that agent over-indexes and drains their account. Transparency is no longer a design choice. It is a security mandate.
The second-order effect is an arms race in platform economics. As agentic AI drives the cost of spam and attacks to near zero, traditional rate limits will fail completely. Founders building consumer platforms must shift their focus from blocking every single attack to bankrupting the attackers. You have to increase the compute or financial friction required to interact with your platform. If it costs an attacker a fraction of a cent to spam your users, your platform will die.
For builders
Make platform abuse financially ruinous
Agents make spamming incredibly cheap. You must build friction that drives up the cost of automated attacks. If you do not, bad actors will overwhelm your infrastructure and you will pay the massive compute bill.
Map tasks by stakes and routine
TikTok uses a two-by-two grid to decide where to deploy AI. Assign agents to routine low-stakes work like initial content moderation. Keep humans on high-stakes highly differentiated tasks where errors cost you money or user trust.
Mandate hands-on AI usage
Chew says the most critical skill for the next three years is using AI tools well. Pay for your team to use tokens. Engineers who do not use tools like OpenClaw cannot understand their security boundaries and will build vulnerable products.
My take
People are building autonomous agents without basic seatbelts. Chew is absolutely right about permissions. If you ask a user for total account access just to automate a simple task, you are building a time bomb. I see too many founders slapping AI lipstick on their apps instead of rethinking the core workflow.
Original reporting: CNA. This is my rewrite and opinion.