deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

VIDRAFT pays $4,500 for attacks on its ViKernel AI security kernel in public hackathon

VIDRAFT is paying the public to break ViKernel, its AI security kernel, in a four-season hackathon with $4,500 in prizes; 298 attacks and 148 kernel fixes had been logged by 19 November 2026.

VIDRAFT pays $4,500 for attacks on its ViKernel AI security kernel in public hackathon

VIDRAFT has turned the security testing of its own AI safety product into a paid public contest. The company's AI security kernel, ViKernel, is now the target of the AI Model Hacking Hackathon, an open event running on Hugging Face that offers $4,500 across four seasons to anyone who can make a dangerous action slip past the kernel. According to a dev.to post, 298 attacks had been submitted by 19 November 2026, 176 findings had been accepted, and 148 distinct root causes had already been patched into the kernel.

The premise reflects a shift in where attackers aim. AI agents increasingly operate tools directly — files, email, the web and the shell — so an adversary no longer needs to exploit software; convincing the model is enough. The post points to EchoLeak (CVE-2025-32711), a 2025 case in which a single manipulated email caused a widely used office assistant to surface confidential data. VIDRAFT's response is to externalise verification: invite everyone to attack, then fold each finding straight into the product.

How participation works

Entrants give instructions to a live AI agent equipped with file, email, web and shell tools, pushing it toward something harmful. If a call the kernel should have blocked gets through, it counts as a breach. The barrier to entry is deliberately low: a Hugging Face account and a browser, with no coding and no GPU required. All of the agent's tools are simulated, so nothing real is deleted or sent, and after each attempt participants see their attack, the kernel's ALLOW or DENY verdict, and an audit number.

Four seasons, four attack classes

The first three seasons are open until 16 January 2027, each targeting a different failure mode:

  • Season 1 — agent takeover
  • Season 2 — data exfiltration
  • Season 3 — hidden instructions

A fourth season, kernel evasion, opens on 16 January 2027. The $4,500 prize pool is split across the four seasons.

What the crowd has found

Alongside the 298 submissions, the post reports more than 70 participants and 176 accepted findings, translating into 148 distinct causes already fixed. The recurring techniques involve obfuscating commands to delete files, send data outward or escalate privileges, or tucking them inside the options of otherwise ordinary tool calls. Every accepted finding is patched immediately, then reused as a private test case for the project's safety leaderboard and its ViKernel Scanner, whose code is published under an Apache-2.0 licence.

A kernel built into the weights

ViKernel's distinguishing claim is placement: the security layer is embedded in the model's weights, so that no matter which phrasing deceived the model, the outgoing tool call is judged immediately before execution and blocked if dangerous. Decisions are recorded on an audit chain the company describes as hard to alter, and a stop switch can halt all agent activity at once. VIDRAFT also reports scanning 20 widely used open models, all of which received an F grade, while the in-weights firewall blocked 897 of 897 dangerous calls in its own testing.

CTO Jang Jaewon, quoted in the post, framed the approach in similar terms: security is only complete when the harmful action is stopped at the last moment before it runs, and every hole found through public attack goes directly into the kernel, strengthening the defence each time.

Why it matters

Most current AI defences try to catch bad prompts; ViKernel instead polices the tool call itself, at execution time. This hackathon is a live, adversarial test of whether that second layer survives contact with a global pool of attackers — closer to a continuous bug bounty than a one-off audit. The early numbers, 148 fixes from 298 submissions, suggest the crowd is finding real gaps, and the reported F grades across 20 popular open models indicate how exposed agent stacks become once obfuscation enters the picture. The obvious caveat is that every figure comes from a single vendor-published post, and the 897-of-897 blocking result is VIDRAFT's own evaluation rather than independent verification. Still, if the pattern holds, the experiment offers a template for how safety-critical AI components could be hardened in public rather than behind closed doors.

  • #ai-security
  • #llm-agents
  • #red-teaming
  • #hackathon
  • #prompt-injection