SShortSingh.
Back to feed

ViKernel 모의해킹 해커톤, 시즌 4개와 총상금 4,500달러로 확대

0
·3 views

허깅페이스에서 11월 17일 시작된 AI 모델 모의해킹 해커톤이 시즌 4개와 총상금 4,500달러 규모로 커졌다. 각 시즌에는 500달러 특별상이 새로 생겼고, 계정당 하루 제출 한도는 10건으로 늘었다. 참가자는 보안 커널 ViKernel이 보호하는 AI 에이전트를 브라우저에서 공격하며, 막아야 할 행동이 허용(ALLOW) 판정을 받으면 점수를 얻는다. 시즌 1~3은 60일간, 시즌 4는 30일간 진행되며 시즌 4에는 최고 난이도 과제가 주어진다. 현재까지 8명이 리더보드에 올랐으며, 상금은 대한민국 법령에 따라 제세공과금을 뗀 뒤 지급된다.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

OpenAI Releases GPT-6.2 Astra and Sol with Lower Token Prices and Migration Caveats

On November 17, 2026, OpenAI launched two frontier models, GPT-6.2 Astra for demanding production workloads and GPT-6.2 Sol for cost-sensitive, high-efficiency tasks. Astra's input and output prices are 20% lower than the previous GPT-6 Astra, and its cached-read rate falls 60%. Sol keeps the same per-token pricing as GPT-6.1 Sol. Developers should not assume a simple model-ID swap will cut costs, since legacy integration logic can drive up token usage and erase the savings. The update also brings a new public beta discount window for the Responses API, alongside long-context surcharges, service-tier multipliers, and a 10% fee on regional endpoints.

0
ProgrammingDEV Community ·

Developer details four rejections and two months getting an MCP server into ChatGPT's plugin directory

A developer behind Pageree, a remote MCP server for building and publishing landing pages with lead capture, has listed it in ChatGPT's plugin directory after several attempts. Submission required a verified publisher identity, a domain challenge file, a short description, test cases, a demo video, and a demo account reviewers could access without signing up. OpenAI's guidelines demand explicit boolean safety hints on every tool, and any tool that overwrites or can reach arbitrary destinations counts as destructive or open-world even if reversible. The developer found that the model ignored a review step in the test case, so they enforced it server-side by rejecting publishes of recently saved drafts. The final review took two weeks, and the developer advises checking everything against the guidelines before each submission.

0
ProgrammingDEV Community ·

Over-Engineered Clean Code Can Slow Teams Down More Than Messy Code

The author recounts a simple form change that took three days and twelve files because the logic was scattered across many layered abstractions. Each class looked clean on its own, yet following the path from a button click to the data required tracing six separate hand-offs. A side-by-side example shows a five-file discount design reduced to a three-line function that performs the same task. The core argument is that the cognitive load of holding many small pieces in mind makes such code costly to maintain, even though no standard tool flags it. The author adds that the ability of AI tools to generate such structures does not remove this burden.

0
ProgrammingDEV Community ·

Production Agent Memory Services Need Separate Metrics for Writes, Retrieval, and Actions

A healthy retrieval dashboard can still hide failures, since a useful record may be dropped before the model call or the task may fail afterward. The article recommends tracking writes, retrieval, context assembly, and final actions as separate stages to locate where problems occur. It cites an Anthropic study from November 6 in which Sonnet 4.6 scored 0.364 on a CL-Bench database task with memory and 0.641 with memory plus refined instructions. Live request data suits reliability and latency measures, while labeled replays or reviewed samples are needed for accuracy-dependent metrics, with sample counts shown beside each rate. The piece proposes specific measures such as write readiness, lookup failures, hit rate at k, context delivery rate, stale context, task success, and memory token usage, each with a defined denominator and data source.