Bug Bounty Architecture for the AI-Slop Era (2026)

Q: Why did curl end its bug bounty program in 2026?

Daniel Stenberg announced that the curl bug bounty would end on 31 January 2026 after running since April 2019, during which it confirmed 87 vulnerabilities and paid out more than $100,000. The confirmation rate had fallen from above 15 percent to below 5 percent in 2025 as AI-generated and poorly researched reports increased, and the stated aim of removing payouts was to remove the incentive for that kind of submission and the load it placed on a small volunteer security team. From 1 February 2026 curl still accepts vulnerability reports through GitHub private vulnerability reporting or by email, with no monetary reward and without a platform intermediary.

Q: What is sloptimism in bug bounty triage?

Sloptimism is the term Bugcrowd's chief AI and science officer David Brumley used in April 2026 for overly optimistic submissions: speculative or AI-generated vulnerability reports filed with minimal or no pre-submission validation and limited context, where the author trusts the language model's output more than the underlying evidence. The reports often look polished, with templated write-ups and confident severity claims, but lack a working proof of concept and frequently describe code paths or endpoints that do not exist. Bugcrowd reported that its triage queues grew by more than 334 percent over three weeks in March 2026 even after excluding legitimate reports.

Q: Should AI be used to triage bug bounty reports?

Yes, for the parts that are checkable, and no, for the decision. A model is cheap and consistent at verifying that a claimed asset is in scope, that a named function or endpoint exists, that no earlier report has the same fingerprint, and at estimating severity from the description. It should never close a report as invalid on its own. Elastic's production system, published in August 2026, runs an eight-stage analysis with an adversarial second pass at roughly $0.50 to $1.15 per report, reproduces around 30 percent of reports in a sandbox at $0.80 to $4.90 each, agrees with human engineers 85 percent of the time on known-outcome reports, and still routes every result to a human before anything is communicated.

Q: How do you stop AI-generated spam reports without losing real researchers?

Raise the cost of submission in ways that a real researcher already pays and a report farm does not: verify identity, tie a reputation tier to how many reports may be in flight, require a structured report with exact reproduction steps, and require a proof of concept that can be executed. Return an incomplete report once with no penalty, so an honest mistake is not punished, but debit reputation for repeated invalid submissions. Bugcrowd's 2026 policies follow this pattern with mandatory identity verification for managed programs, throttling of low-performing accounts, suspensions after ten consecutive invalid automated reports and permanent bans for submission farming. None of these measures require reading the report.

Q: What does a bug bounty reproduction harness look like?

A reproduction harness provisions an ephemeral, isolated instance of the in-scope asset at the reported version, executes the submitted proof-of-concept steps exactly as written, captures exit codes, HTTP traces, logs and screenshots into an evidence bundle keyed by the report, and destroys the sandbox afterwards. It must not repair a broken exploit, because the goal is to test the submitter's claim rather than to find a bug nearby. It must not hold production data, long-lived credentials or general network egress, because the proof of concept is code written by an untrusted party. Reports that reproduce go to a human with the evidence attached; reports that do not are returned once with the harness output.

Q: Should a bounty pay on report acceptance or on reproduced impact?

Pay on reproduced impact. In 2023 a write-up was a reasonable proxy for effort, because writing one required understanding the bug. In 2026 a plausible write-up costs minutes, so paying for prose pays for the thing that became free. Attaching the payout to a reproduction in your own sandbox, with severity set by exploitability rather than by the researcher's claim, aligns the incentive with the only step a language model cannot fake. It also lets you pay faster, because a reproduced finding with an evidence bundle is a fact that finance and legal can see rather than a judgement they have to trust.

Q: When should a company run a vulnerability disclosure program instead of a paid bounty?

Run a disclosure program without payout when you cannot fund the validation side: no capacity to verify identities, throttle accounts, maintain sandbox images for in-scope assets or reproduce reports quickly. A paid bounty without those controls attracts generated volume you cannot process, which is what happened to curl. A disclosure program with a structured intake, a reproduction requirement and a stated response window still gives researchers a safe route and gives you the findings, and the money goes into fixing what arrives. Start private and invited, with a scoped asset list that has an owner and a sandbox image, and add payouts only when the reproduction rate for that scope is stable.

Q: How does the EU Cyber Resilience Act affect bug bounty programs?

The Cyber Resilience Act imposes vulnerability-handling obligations on manufacturers of products with digital elements, including having a coordinated disclosure policy, a contact point for reporting, and processes to address and remediate vulnerabilities, with reporting duties for actively exploited vulnerabilities. For a bounty program this means a reproduced report is a regulatory event, not just a payout decision: the clock for handling and, where applicable, reporting starts when the vulnerability is confirmed, so the reproduction harness and the evidence ledger become part of the compliance record. Programs selling into the EU should route reproduced findings into the same vulnerability-management queue that tracks those obligations.

Satyam Kumar

ブログに戻る

cyber-security-patterns

Bug Bounty in the AI-Slop Era: Pay for the Reproduction, Not the Report (2026)

By Satyam KumarSeptember 24, 202614 min read

bug bounty program vulnerability disclosure program ai slop sloptimism bug bounty triage hackerone bugcrowd curl bug bounty proof of concept reproduction security triage automation vulnerability management appsec program responsible disclosure llm generated reports security validation sandbox cyber resilience act vulnerability handling

Bug Bounty in the AI-Slop Era: Pay for the Reproduction, Not the Report (2026)

Frequently Asked Questions

この記事を共有する

Twitter LinkedIn WhatsApp

Satyam Kumar

Founder & AI Architect, AppScale LLP

AI＆クラウドアーキテクト。数百万人にスケールするシステム構築を支援。

LinkedIn GitHub

Bug Bounty in the AI-Slop Era: Pay for the Reproduction, Not the Report (2026)

Frequently Asked Questions

この記事を共有する

Comments

Leave a comment