curl project

The maintainers of large open source projects are taking the worst of the AI slop. In a Linux Foundation survey of 234 open source contributors and maintainers, people running projects with more than 50 contributors estimated that close to a third of the code changes landing in their queue are low quality or AI-generated. Those maintainers now worry more about burning out than about the code itself. Across all projects, the estimate comes to about … More → The post AI slop is piling up highest on the biggest open-source projects appeared first on Help Net Security .

AI is disruptive. Anthropic’s Claude Mythos model, and its successors, promise to be even more disruptive: they could threaten the existing bug bounty and/or in-house offensive security industries. AI has been widely adopted by both cybersecurity attackers and defenders. Attackers use it to help find bugs and craft attacks from sophisticated social engineering through to developing exploit and malware code. Defenders use it to help detect attacks in progress, detect deepfakes, and help code new software, and for bug bounty hunters and offensive security practitioners, to unearth bugs to fix them before they can be exploited. So far, AI has proven to be a force multiplier rather than a position replacement. Mythos threatens to alter this balance. The evolution of bug bounty programs Bug bounties and pentesting are in a state of flux. That’s nothing new: everything in cybersecurity is constantly in flux. But the Mythos arrival may provide the most rapid flux in offensive security yet. A bounty is a reward. ‘Dead or alive’, was an early 19th-century US tagline. That concept still survives, but with law enforcement now offering bounties for information on live cybercriminals. A bug bounty is a reward for finding a bug not a person. In 1983, Hunter & Ready offered a free Volkswagen Beetle car (commonly known as a Bug) as the reward for finding a computer bug in its VRTX operating system. The new tagline was ‘Get a bug if you find a bug’. The concept of bug bounties had arrived and began to expand from the 1990s: Netscape in 1995; IDefence introducing the middleman concept in 2002 (any person could report any bug to any vendor); Mozilla for Firefox in 2004, Google in 2010, and Facebook in 2011. The HackerOne (with Kara Sprague as CEO) and Bugcrowd (co-founded by Casey Ellis) bug-bounty platforms were established in 2012, followed by YesWeHack in 2015, and Intigriti (Inti De Ceukelaire) in 2016. These are the four primary bounty platforms. Throughout the 2010s the concept expanded and many more companies began to offer bug bounties. By 2022, bounty hunter Youssef Samouda was able to tell SecurityWeek, “With Meta and Google, I make around $400,000 per year.” At the end of 2022, AI in the form of LLMs became generally available, and by late 2024 and early 2025 the modern concept of autonomous agentic AI began to take center stage. By June 2025, autonomous offensive security firm XBOW, had achieved #1 position in HackerOne’s leaderboard. The history of bug bounties shows a consistent combination of expansion with an increasing use of automation and artificial intelligence – which brings us to today. In-house offensive security has followed a similar path but driven by salary rather than reward. Bug bounty today Cassim Khouani (known online as Aituglo and listed in the top 30 Hackers on YesWeHack) wrote The state of Bug Bounty in 2026, published on April 13, 2026. In it, he describes using Claude to aid discovery. Overnight, it discovered ten bugs. “Sounds great on paper. Except half of them were duplicates, and the rest took weeks to get triaged because the report queue on that program had become unmanageable. Welcome to bug bounty in 2026.” He believes bug bounty as we know it today is dying. “What comes next can be better, if we play it right.” Everybody, he suggests, is using one or other form of AI to search for bugs, 24/7 without getting tired. It succeeds, but with side effects: “We end up in a constant mental fog, jumping from one tmux pane to another, switching from one program to the next.” And the bounty platforms themselves suffer from so many new AI-assisted submissions, with triaging and payments taking longer. Companies paying bounties are also suffering with more bug reports, some of poor quality and some critical. “More and more companies are stepping back from bug bounty,” he writes, “while others [such as Google] increase their rewards or change their policy.” Note that after Khouani wrote this in mid-April 2026, Google lowered its Chrome bug bounties and raised its Android bounties on April 30, 2026, citing AI as the cause for both. This is flux with its foot on the pedal: rapid change but not necessarily for the better. “The bug bounty of 2024 is dead. The one in 2026 is a different sport. The hunters who will make it are not those who launch the most agents, but those who know what to look for and where to look. AI is a multiplier, not a replacement.” This was written by Khouani based on his experience of using Claude to assist in bug finding. But then along came Claude Mythos, announced almost at the same moment he published his article. Anthropic’s claims for Mythos going forward suggest the future of AI is more than just a force multiplier. Mythos discovering vulnerabilities Mythos reportedly performs better than any other AI model in finding zero day bugs. “Over the past few weeks, we have used Claude Mythos Preview to identify thousands of zero-day vulnerabilities (that is, flaws that were previously unknown to the software’s developers),” announced Anthropic on April 7, 2026, “many of them critical, in every major operating system and every major web browser, along with a range of other important pieces of software.” In May 2026, Anthropic said its Mythos Preview had identified more than 23,000 potential vulnerabilities after scanning thousands of open-source software projects. Anthropic is apparently so concerned about its ability to find unfound bugs that it has released Mythos Preview to major software providers, allowing them to find and fix their own vulnerabilities (Project Glasswing) before the model becomes generally available. The CSA is equally concerned, having published a paper titled, “The ‘AI Vulnerability Storm’: Building a ‘Mythos-ready’ Security Program”, in which it recommends: “Introduce AI agents to the cyber workforce across the board, enabling defenders to match attackers speed and begin closing the gap.” Fed Chair Powell and Treasury Secretary Bessent met with the heads of major US banks to discuss the cyber risks that may be introduced by Mythos; Reuters has reported ‘Banking industry scrambles for Anthropic’s Mythos as global regulators review risks’; and the media has been full of wild, weird, and wonderful reporting. But one area has had little reporting so far: if Mythos is capable of finding bugs and developing exploit chains so rapidly, what effect will this have on the value and future of the existing bug bounty and offensive security industries? Organizations could just point Mythos at their software and find the bugs without needing to pay bounties or employ expensive pentesters and red teams. The future of bug bounty and offensive security Bug bounty and offensive security are not going away; but both must adapt to a new reality. AI is like sniping: the projectile and its effect may be autonomous, but it still needs a human to aim and pull the trigger. Complete autonomy is still in the future, and that human involvement will remain for years to come. It’s the speed of delivery and the accuracy that has changed. Keep Calm and Carry On: Mythos is not revolutionary Tod Beardsley, VP of security research at runZero, counsels that Mythos should be viewed in the historical concept of an industry barely 30 years old: any advance will seem huge and disruptive while it’s happening. “To be blunt, I don’t think Mythos is fundamentally different or the ‘YOU MUST BUY THIS’ security tooling that Anthropic’s marketing would like us to believe. It’s better tooling, for sure…” But, he adds, “To think that this (or any) model is so fundamentally powerful, dangerous, and revolutionary that you’d be a fool to not buy is to ride along with the classic FUD-based marketing that so often colors cybersecurity marketing… This is just another step on the road to better understanding the risk profile of your particular network.” Richard Ford, CTO at Integrity360, agrees with the need to stay calm but adds, ‘be ready to adapt’. “Anthropic’s own system card shows that this level of performance relied on uncensored models, extended compute, and heavy resampling. In other words, this is not yet a real-world scenario.” Nevertheless, he adds, “Bug bounty programs and human-led testing rely on expertise and time. AI will start to reshape that, although areas like business logic will still depend on human understanding.” It’s adaptation, not replacement, says Chris Payne, VP of forward deployed cyber engineers at Sevii. “Discovery accelerates for everyone, but the real bottleneck has always been investigation and remediation. The defenders who win will pair agentic AI with strong governance so they can investigate, hunt, and remediate at machine speed and endless scale, which will close the gap attackers are widening as we speak.” Jon David, co-founder and MD at NR Labs, agrees with this. The power of Mythos and future AI will allow attackers to find and exploit vulnerabilities faster than defenders can fix them. “But if we leverage AI in the same capacity, it also allows us to find and patch the vulnerabilities before they’re public in production. We must just make sure we’re using the same capabilities as the attackers on ourselves before they do.” This should ensure the continuance of in-house offensive security teams. Bounty hunters and bounty platforms have a slightly different problem to solve: participating companies are likely to become reluctant to pay bounties for an increasing volume of existing but largely irrelevant bug slop already found by hunters and their AI agents, and likely to increase with Mythos. “The widely used Curl project ended its HackerOne program in January 2026 because over 95% of submissions were AI-generated junk,” comments Melissa Bischoping, head of threat research & intelligence at Tanium. “HackerOne [also] paused the Internet Bug Bounty in March, explicitly citing an imbalance between AI-assisted discovery and remediation capacity. Mythos makes this worse by an order of magnitude.” The need for adaptation. Evolving AI increases the speed of discovery and decreases the time to exploitation. Kara Sprague, CEO at HackerOne, points out this is not the gap meant in the CSA report. “The gap they are describing isn’t the gap from discovery to exploitation; it’s clear that it has already collapsed. The gap they are referring to is the operational gap between discovery and remediation: organizations still patch on human timelines, manage risk through quarterly assessments and run remediation through change processes that were never designed for AI-velocity threats.” AI-assisted discovery has flooded programs with low-to-mid-severity findings that maintainers cannot absorb. This will increase with Mythos. “The incentives of such bounty programs need to be rebalanced to favor remediation, which is now the key constraint,” she continues. “Vulnerability findings often sit for a long time before remediation, and this trend will continue as the volume increases. It then becomes a question of prioritizing learning from mistakes rather than focusing on individual issues as that approach hasn’t been scalable and certainly won’t be as these advanced models become more distributed,” adds Shlomie Liberow, founder at aisy.ai, and formerly head of hacker R&D at HackerOne. Bounty platforms are already strained by the number of bugs being discovered. And corporations are unable to keep pace with existing patch levels. What is required is the ability to prioritize high severity bugs over low value bugs, and incentives to prioritize high severity over low value will be built into the programs. The longer view. Corporate options for vulnerability detection remain primarily third-party bounty hunters or in-house offensive security teams. The best option will be governed by whichever adopts the latest technology functionality faster and more intelligently. “Today, frontier model providers (like Anthropic) make AI abundantly cheap to use, reducing the cost of vulnerability research,” suggests Aaron Sant Miller, VP and AI lead for Booz Allen’s Integrated Cyber Business. “If organizations increasingly choose to bring this function in-house, you can expect the bug bounty industry to take a hit. Conversely, bounty hunters may find AI reduces their own cost and time – we may see more bounties executed per month, at a lower cost per bounty. Today’s balancing act will be shaped by adoption.” It may be different tomorrow. The current cost of AI is being eaten by the frontier model developers, and the price to users is artificially low. Once the market has been hooked, the developers may increase their prices to reflect the true market cost. When that happens, organizations will reconsider outsourcing again. “Once AI usage costs begin to reflect their true operating costs, the cost of bug discovery will normalize; organizations and hunters alike will have to determine the fair market price for vulnerability discovery,” he continues. “Overall, in-house security teams that readily adopt AI and evolve their workflows will be more resilient to the disruption.” Sprague agrees that the market for vulnerability hunting is already changing, but she notes, “The total value of the bounty market is growing, not shrinking. High-severity, business-logic and AI-specific vulnerability research (for example, prompt injection, model extraction, adversarial manipulation) is paying more than ever, because very few researchers can do it well.” For in-house offensive teams, she says, “Red teaming was already evolving beyond ‘can we get in’ toward mapping business process fraud paths, executive targeting, third-party compromise chains, privileged identity abuse across SaaS estates. Mythos can’t model your business. It can’t tell you that the real crown jewel isn’t in the codebase at all.” She sees continuous AI-executed adversarial testing with human expertise woven throughout. “Novel attack paths, business logic vulnerabilities and the creative lateral thinking that turns a low-severity finding into a crown jewel compromise still require human ingenuity. AI raises the floor; it doesn’t replace the ceiling.” Summary AI is changing the rules for human vulnerability hunters, but not replacing the requirement. AI can find a never-ending volume of bugs – they will always exist, in both old and new software. But this new AI-discovered volume of bugs, with many of them being low severity, ultimately inconsequential slop, has highlighted the real problem. It’s not finding the bugs (that hasn’t been a problem for years); it’s distinguishing the meat from the slop with rapid remediation. AI is not good at this. It requires human knowledge and ingenuity. In-house human offensive security, using AI but not relying on it, can do this well – and the value of in-house detection and remediation will remain. External bounty hunting will be most impacted, but will continue to offer a valuable service if the bounty paid can be refocused on critical bugs and remediation. This will offer a valuable alternative for those companies that find employment of a permanent in-house team to be too costly. Vulnerability hunting, like every other aspect of cybersecurity, is subject to new technology. And like every other aspect of cybersecurity, the advice to practitioners remains the same: keep calm, adapt, and carry on. Related: Cyber Insights 2026: Offensive Security; Where It Is and Where It’s Going Related: OpenAI Launches Bug Bounty Program for Abuse and Safety Risks Related: Microsoft Bug Bounty Program Expanded to Third-Party Code

AI is drowning software maintainers in junk security reports AI-assisted vulnerability research has exploded, unleashing a firehose of low-quality reports on overworked software maintainers who are wasting hours sifting through noise instead of fixing real problems. Linus Torvalds, the Linux kernel’s creator, says the flood has made the project’s security mailing list “almost entirely unmanageable, with enormous duplication due to different people finding the same things with the same tools.” Too many duplicates, and too much AI slop “If you found a bug using AI tools, the chances are somebody else found it too,” Torvalds wrote in the note accompanying the latest Linux kernel release candidate. “If you actually want to add value, read the documentation, create a patch too, and add some real value on *top* of what the AI did. Don’t be the drive-by ‘send a random report with no real understanding’ kind of person.” Jarom Brown, Senior Product Security Engineer at GitHub, acknowledged last week that while AI lowering the barrier to entry for security research is a welcome development, his team is being inundated by submissions that fail to demonstrate any real security impact. These include reports without a proof of concept, theoretical attack scenarios that don’t hold up under scrutiny, and findings already covered by GitHub’s published ineligible list. And GitHub isn’t the only recipient of this unwanted deluge. “Programs across the industry are grappling with the same challenge, and some have shut down entirely,” he said. GitHub has stopped short of such a drastic measure, but is now requiring submitters to validate AI-assisted findings before sending them in. Going forward, a complete submission must also include a working proof of concept demonstrating exploitation potential and concrete security impact. Also, reports covering known ineligible categories will be closed as Not Applicable, which may impact the submitter’s HackerOne Signal and reputation, he added. Finally, Brown urged researchers to be concise: bloated, AI-padded reports slow down triage and waste everyone’s time. The researcher’s view The collateral damage extends beyond the programs themselves. Shubham Shah, co-founder of Assetnote and a respected security researcher, says organizations are now taking far longer to review legitimate reports and act on real flaws, and that’s killing the feedback loop that keeps top researchers engaged. While bug bounty platforms like HackerOne and Bugcrowd are trying to fight the onslaught of AI-created spam reports with AI and added controls, he says that “the joy of reporting vulnerabilities to bug bounties is quickly dissipating” – and not just for him. “Hopefully the platforms actually work this out, but until then, I can’t see myself continuing to report high quality original research to certain programs where I have meaningfully contributed for a decade when they fail to understand the difference between myself and a researcher that doesn’t have any credibility,” he added. In the near term, some experienced researchers may retreat to private vulnerability research and invite-only bounties. Open source bears the brunt The AI-powered “industrialization” of vulnerability discovery is currently a much bigger problem for open source projects than big organizations like Microsoft or Google, as they rely on volunteer maintainers, whose number and time is limited. Those limitations have, for example, led the cURL project to stop accepting HackerOne submissions and eliminate monetary rewards for security reports. CURL lead developer Daniel Stenberg hoped the latter decision would remove the incentive for submitting AI slop, and stated his belief that “the best and our most valued security reporters still will tell us when they find security vulnerabilities.” The project switched to welcoming reports via GitHub or email, but a month later reverted to using HackerOne because those two avenues proved less effective than expected for reporting vulnerabilities. However, the project decided to stick with their decision not to offer bounties for bug reports. “From that day, the nature of the security report submissions have changed. The slop situation is not a problem anymore,” Stenberg noted in April. The number of reports rose, their quality was higher (even if they were compiled with the help of AI), and the rate of confirmed vulnerabilities surpassed the 2024 pre-AI level. While that change was welcome, Stenberg believes that the raised influx of “good” vulnerability reports will present a different problem for open source projects. “This avalanche is going to make maintainer overload even worse. Some projects will have a hard time to handle this kind of backlog expansion without any added maintainers to help,” he pointed out. In the wake of cURL’s departure and return, HackerOne acknowledged the problem AI slop may represent for under-resourced organizations, and advised customers to refine the scope and submission guidelines to reduce noise, use AI-assisted triage tools, and pair that automation with human oversight. “As AI makes it easier to automate submissions, preserving signal quality becomes critical so open source maintainers can stay focused on fixing real issues,” Michiel Prins, Co-founder & Senior Director, Product Management at HackerOne, told Help Net Security. “Our focus is helping programs manage that shift with workflows that filter noise early, surface credible reports, and keep vulnerability management sustainable, so open source communities can maintain the transparency and resilience they’re known for” The Open Source Security Foundation’s Vulnerability Disclosures Working Group is also seeking community feedback as it works to help open source maintainers tackle AI-generated junk reports. Its goals include compiling best practices, creating policy templates, and developing guidance to help maintainers spot and handle AI-assisted submissions. UPDATE (May 19, 2026, 06:55 a.m. ET): This article has been updated to reflect another change in cURL project’s bug bounty program, which happened in February 2026. The original article stated that the program was moved from HackerOne to GitHub, but we were unaware that cURL returned to HackerOne since then. The article has also been updated to include additional findings that emerged as a result of this switch and the elimination of rewards for vulnerability reports. Subscribe to our breaking news e-mail alert to never miss out on the latest breaches, vulnerabilities and cybersecurity threats. Subscribe here!

The developer of the popular curl command-line utility and library announced that the project will end its HackerOne security bug bounty program at the end of this month, after being overwhelmed by low-quality AI-generated vulnerability reports. The change was first discovered in a pending commit to curl's BUG-BOUNTY.md documentation, which removes all references to the HackerOne program. Once merged, the file will be updated to state that the curl project no longer offers any rewards for reported bugs or vulnerabilities and will not help researchers obtain compensation from third parties either. "Up until the end of January 2026 there was a curl bug bounty. It is no more. The curl project no longer offers any rewards for reported bugs or vulnerabilities. We also do not aid security researchers to get such rewards for curl problems from other sources either," reads the upcoming update. curl is a command-line utility that allows you to transfer data over various protocols, most commonly used to connect to websites. An associated libcurl library enables developers to incorporate curl into their applications for easy file transfer support. Since 2019, its bug bounty program has been run through HackerOne and the Internet Bug Bounty, offering cash rewards for responsibly disclosed security vulnerabilities in curl and libcurl. Daniel Stenberg, curl's founder and lead developer, says the program has seen a large increase in low-effort and invalid reports, many of which appear to be AI-generated slop. AI slop is the growing flood of low-effort, AI-generated content that sounds good but doesn't actually contain anything useful or productive. In a recent post to his personal mailing list, Stenberg explains that these low-quality reports are straining the curl security team, leading him to withdraw from the program. "We started out the week receiving seven Hackerone issues within a sixteen hour period. Some of them were true and proper bugs, and taking care of this lot took a good while. Eventually we concluded that none of them identified a vulnerability and we now count twenty submissions done already in 2026," explained Stenberg. "The main goal with shutting down the bounty is to remove the incentive for people to submit crap and non-well researched reports to us. AI generated or not. The current torrent of submissions put a high load on the curl security team and this is an attempt to reduce the noise," continued his post. In comments on the pull request, Stenberg said that withdrawing from HackerOne may not stop the flood of junk reports. However, he said that curl is a small open-source project with a limited number of active maintainers, and that, to ensure its survival and protect developers' mental health, he needed to take this action. Stenberg has also shared examples of what he considers AI slop reports and said he has seen a steep rise in security submissions at curl compared to other open-source projects. "We seem to have data that confirms that the #curl bug-bounty has received a steep increased submission rate through 2025, while several other Open Source programs also hosted on Hackerone have not," Stenberg posted to Mastodon. The switch from HackerOne's bug bounty program to an internal submission process will happen in stages. Stenberg says the curl project will accept HackerOne submissions until January 31, 2026, and that any reports in progress at that time will continue to be processed. Starting February 1, 2026, the project will no longer accept new HackerOne submissions and will instead ask researchers to report security issues directly through GitHub. Curl's new stance is also reflected in a recent update to its security.txt file, which states that the project offers no monetary compensation for reported vulnerabilities and warns that people who submit "crap" reports will be banned and ridiculed publicly. Stenberg says he will publish a blog post next week with more details about this upcoming change. After publishing our article, HackerOne told BleepingComputer that they are seeing increased volume in AI-helped reports across the entire security industry. The company says they have introduced AI features to help companies filter low-quality reports and filter duplicates. “The pressure of added volume can feel acute in open source, where small teams and volunteers often shoulder vulnerability intake and validation directly, without extra buffering layers," explained HackerOne. "curl’s experience is not representative of what we typically see across enterprise engagements. Most organizations rely on HackerOne’s managed triage and AI workflows so internal teams stay focused on findings that warrant deeper investigation." "With the right architecture and support, bug bounties remain one of the most powerful tools for uncovering novel, hard-to-find exposures that drive real-world risk." Update 1/28/26: Added a HackerOne statement. Overall prevention scores can hide what happens after initial access. Once attackers are using valid credentials, prevention drops sharply. The Blue Report 2026 measures defenses technique by technique across 338 million simulations run in customer production environments. Get the report