Breaking

PWN.AI Says Its Agent Captured Google kvmCTF Flag With KVM Use-After-Free Exploit

September 28, 2026 at 17:57 EDT

AI Agents
Security

PWN.AI said its autonomous security-research system captured a Google kvmCTF flag by driving a KVM x86 use-after-free exploit from inside an authorized guest VM. The company says the June 17 live run forced Google's KASAN-backed challenge host to return the 64-bit token cdc63bb4b015eb8c, after three official attempts returned zero.

June 17, 2026 · PWN.AI

Agent reaches a kvmCTF flag

PWN.AI says its security-research system drove a KVM x86 use-after-free chain until Google's challenge host returned a 64-bit read flag. The run was authorized, narrow, and technically demanding.

14,338

harness lines

49,152

sparse mappings

1,002

recycle aliases

official attempts

Harness growth, shown to scale

Each block represents roughly 1,000 retained lines of exploit harness.

3,304

prototype

12,955

private writer

14,338

winning run

Exploit path

VMREAD retarget
EPT12 entry is rewritten without page-track cleanup

→

Stale rmap
A reverse-map pointer survives after shadow memory is freed

→

HC102 flag
The host records a qualifying invalid read and returns the token

Why it matters

The campaign connects source review, allocator tuning, live repair, and an official challenge flag rather than stopping at a plausible crash.

Key limits

The target was authorized, humans set scope and infrastructure, the system was a specialized harness, and no complete model transcript was retained.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.