Science & Tech · Cybersecurity · 1 hr ago
Vercel Sandbox flaw could let AI agents leak sensitive data, researchers say
Researchers found a flaw in Vercel Sandbox that could let rogue AI agents escape its protections and leak sensitive information.
The tool runs code in isolated environments, and a certificate-handling weakness could bypass network restrictions meant to block connections to unapproved destinations.
Sandboxes are used to limit what AI agents can access and where they can send data.
Researchers have found no evidence of exploitation, but they cannot rule out that the flaw has been used.
Vercel told the researchers it needs time to work out how to patch the problem.
An approved server could unknowingly relay stolen data, showing that security checks must consider what happens after an initial connection.
Researchers say a flaw in Vercel Sandbox could let rogue AI agents bypass network restrictions and transmit sensitive information.
The flaw involves certificate validation and could let an approved server relay data to an attacker-controlled server.
The researchers, from AI security startup AIR, said they found no evidence the flaw has been exploited, but could not rule out exploitation.
AIR said it contacted Vercel, which said it needed time to find a patch.
- Who
- Researchers at AI security startup AIR found the flaw. Vercel makes the affected sandbox.
- What
- A network-filtering flaw could let rogue AI agents bypass restrictions and leak sensitive data.
- When
- The research was published on October 13, 2026. The researchers said they discovered the flaw on Monday.
- Where
- Vercel Sandbox, a tool for running code in isolated environments.
- Why
- The flaw in certificate handling could allow an approved server to trigger a request to an attacker-controlled server, bypassing network restrictions.
This story does not have two clearly opposing sides.
Researchers said they discovered a vulnerability in Vercel Sandbox.
AIR contacted Vercel about the vulnerability, and the company said it needed time to find a patch.
The research was published.
- Affected tool
- Vercel Sandbox
- Researchers
- AI security startup AIR
- Vulnerability
- Network filtering flaw involving digital certificate handling
- Potential impact
- Sensitive information could be transmitted to an attacker-controlled server
- Exploitation evidence
- AIR said it had found no evidence of exploitation in the wild, but could not rule it out
What people are saying
This vulnerability shows that enterprises can have Hugging Face [style] mini incidents with their own agents
It’s a very hard leak to patch
A crafted certificate is all it takes
Security has to account for what happens after the first connection.
Sources
Related news
Cyberattack disrupts IT systems at three TFG London fashion brands
Canadian cybersecurity executive arrested in ShinyHunters investigation
Cyberattack disrupts refrigeration at hundreds of French supermarkets
Norway fines Microsoft-owned Xandr $9.6 million over Grindr data sharing
GitLab fixes maximum-severity flaw that could let attackers run server code
Ledger links CryptoBilis wallet implants to confirmed crypto thefts
DDoS attack disrupts Canonical websites for more than four hours
GitHub adds 48-hour limit after npm publishing flaw exposed 1,500 packages
Survey finds human error is the leading cybersecurity concern
Ahold Delhaize takes systems offline after cybersecurity incident
Cyberattacks target refrigeration systems at more French supermarkets
