What Happened to HackerOne?

So…what’s going connected astatine HackerOne lately? It mightiness beryllium clip for a wellness check.

If you are caller to the bug bounty abstraction (1-3 years), you mightiness not person immoderate thought what I’m talking about.

But arsenic a decently washed-up bug bounty huntsman who lived done the aureate era of HackerOne, I deliberation it’s clip to reside the elephant successful the room.

For immoderate context, I started arsenic a hacker connected HackerOne successful 2017. When I began moving successful tech, that hands-on acquisition was highly useful for managing a bug bounty program, since I knew what researchers wanted, and really to interact pinch them.

As a result, I person managed aggregate ample bug bounty programs connected HackerOne crossed various companies from 2018 to 2025 and I’ve been connected some sides of the equation.

What I’m astir to talk astir comes from first-hand experience, some arsenic a interrogator and arsenic a bug bounty programme manager, and many, galore years of nonstop conversations pinch HackerOne, some publically and privately.

Background #

To start, I deliberation it’s important to recognize what HackerOne was primitively designed to be.

In 2011, 2 ethical hackers, Jobert Abma and Michiel Prins, group retired to find information vulnerabilities successful 100 of the largest tech companies. They succeeded and recovered bugs successful Google, Facebook, Apple, Microsoft, Twitter, and galore others. At this constituent successful time, the scenery for ethical information investigation was risky, legally dubious, and very scary for information researchers.

Not only was location important individual liability, but location had been aggregate instances of hackers being criminally charged and sentenced to jailhouse clip for uncovering and reporting information vulnerabilities anterior to this. Much of this was owed to circumstantial arbitrary lines drawn successful the soil which, if crossed, made you a bad actor, but if not crossed, made you a perchance bad character but technically not one.

Bug Bounty Platforms for illustration HackerOne were designed to straight reside this issue. It created a safe communal abstraction for companies and hackers to connect, and it paved the measurement for ethical hackers to taxable information vulnerabilities to companies, pinch afloat consent, and get paid for that work. This was a immense milestone. You nary longer had to interest astir getting dragged to tribunal (or jail) for uncovering an IDOR that leaked customer data. Instead, you sewage a “thank you” and a rate payout for making everyone a small safer.

This operating exemplary was the instauration for bug bounty and remained that measurement for the adjacent 5+ years.

The Golden Age of HackerOne and Live Hacking Events #

During this period, location was a very beardown and definitive attraction for the business: really do we make this the champion imaginable merchandise for hackers?

The group moving the business day-to-day were hackers, hacker-adjacent, and astir (if not all) were face-to-face pinch hackers connected a regular basis.

From 2017 to 2020, HackerOne was doing Live Hacking Events (LHEs) each fewer months. These were exclusive events wherever the apical bug bounty researchers astir the world would alert into a location, beryllium fixed a target, and spell perfectly ham uncovering captious vulnerabilities. LHEs were a immense worth prop for programs. During a 1-3 time period, you would get much precocious and captious information reports than you would person received for the full twelvemonth otherwise.

Every arena had a 1-of-1 civilization designed poster pinch graphics, hacker usernames, stickers and situation coins. It’s difficult to overstate what an unthinkable and productive play this was for HackerOne and their apical programs.

These events were exclusive and highly coveted; invites and +1s were practically their ain currency. And the situation astatine these events was surreal. You would beryllium fixed free flights and hotels astir the world, and walk a fewer days surrounded by the champion and astir skilled bug bounty hunters successful the world. These researchers would regularly find immoderate of the astir impactful bugs utilizing their ain caller techniques, and each while sharing tips and tricks successful one-off conversations that could not beryllium replicated anyplace else.

Prior to the advent of unrecorded hacking events, astir information interrogator circles were small, isolated, and sharing accusation publically was practically unheard of. LHEs created a measurement for information researchers to link pinch each other, and fundamentally created a full caller organization wrong infosec. Most of my closest friends nowadays are group who I met done the unrecorded hacking scene, and I americium highly grateful to HackerOne for that.

LHEs were not the only area wherever HackerOne was building and found a organization for information researchers. They created a HackerOne Community abstraction to shape meetups, online events, workshops, and CTFs. They created location clubs, and appointed hackers who lived location arsenic ambassadors to thief foster and turn section interrogator communities each astir the world.

But slow but surely, things began to change. The organization groups and events mislaid momentum and fizzled out. The civilization designed silkscreen LHE posters became inexpensive low-effort laser prints. The group who had dedicated years to creating and moving unthinkable events were laid disconnected aliases left. The LHE invitation and scoring systems became (even more) exclusive, gamified, and exceedingly calculated.

So 1 by one, the dominoes began to fall.

The Profit Problem #

Before going further, I deliberation it’s important to get into the “why” down these changes that started happening. Sometime astir 2020 aliases 2021, HackerOne was forced to travel face-to-face pinch a very important mobility that each business must inquire itself astatine immoderate point: “How do we make money?”

To understand really HackerOne moreover was capable to past arsenic a business, you person to first cognize that HackerOne was fundamentally moving wholly connected VC money for the first 10 years of its existence. Between 2014 and 2022, the institution did a seed information almost each 2 years, raising a full of $160M. If a institution is self-sustaining, location is little-to-no logic to proceed raising money unless you person an incredibly precocious burn-rate, which would beryllium overseas for a institution pinch specified small infrastructure and method innovation.

If you cognize thing astir VC, you cognize really this tends to work; they springiness you money, and successful return, you springiness them indirect power of the business done committee seats, advisory positions, and different leverage mechanisms. The VCs gave the money, truthful the VCs get the power.

Rome did not autumn successful a day, and neither did HackerOne. The paradigm displacement started slowly, and past each astatine once. The halfway exertion driving the level began to stagnate. Very small changed wrong the platform. The UI remained tired and “functional”. The capacity remained lacking.

But the VCs looked astatine their playbook and realized that the easiest measurement to make money was simple: get much customers.

First, the founding CEO was replaced pinch a firm CEO. Instead of a 20% trim of bounties paid, they shifted to capacity-based interest structures, yearly contracts, and locking customers into multi-year deals.

And then, sales. SALES SALES SALES!! They were afloat convinced that income were the breadstuff and food of the business. So alternatively of basing the institution astir hackers, they leaned into sales.

More customers = much yearly contracts = much predictable semipermanent revenue. Once a customer is locked in, they are fed stats and numbers and greased up to support them happy, but besides to support demands low. When statement renewal comes up, they are fixed monolithic (30-60%) discounts successful speech for a multi-year statement successful bid to support them locked successful and paying.

Account Managers would person regular meetings pinch customers and promote them to raise bounties to enactment competitory and charismatic to hackers. They would show you, “Hackers want to walk clip focused connected the highest-paying programs”.

HackerOne was proud of this and encouraged the behaviour internally. They gave retired rewards, free vacations, and encouraged their quickly increasing income squad to motion caller customers arsenic overmuch arsenic possible.

As this continued, things continued to degrade for everyone involved.

For triagers, the champion ones (who were hackers themselves) burned retired and quit. The salary was excessively debased and the workload was excessively much.

For hackers, the triage acquisition declined, and they became flooded pinch excessive amounts of low-quality programs.

For customers, the value of reports went down, the costs went up, and the title to the bottommost began.

And for everyone, the level acquisition declined and ne'er evolved.

Typically astatine this point, marketplace economics would dictate that this is wherever a competitor takes your business and you must amended to clasp marketplace share. But bug bounty is an oligopoly, and location are 3 companies that power almost the full market.

So instead, HackerOne was forced to create a caller exclusive opportunity called the “Hacker Success Program” (HSP). This is simply a dedicated abstraction for apical hackers to get nonstop support from HackerOne for thing bug bounty related. They were assigned to a Hacker Success Manager (HSM) who was employed by HackerOne, and they could leverage that narration to navigate miscommunications, bad programme experiences, debased aliases inaccurate bounty payouts, and overmuch more. They besides usage this abstraction to connection unsocial opportunities for illustration H1 challenges for hackers successful the HSP group.

One point I want to telephone retired is that while this is simply a awesome thought successful theory, successful believe it created a wholly lopsided playing section for caller hackers. You person fundamentally zero ways to advocator aliases activity done problems (good luck moving pinch HackerOne support), truthful the rich | get richer and you are fixed the action to woody pinch it aliases get mislaid unless/until you are a apical hacker. Oddly enough, it would beryllium overmuch easier to go a apical hacker if you had entree to these resources successful the first place.

Anyway, for immoderate hackers, the HSP revolutionized what was antecedently an intolerable ceramic wall. If you sewage screwed complete by a institution who grounded to understand the information effect of your report, you could now thin connected your HSM to thief get a nonstop statement of connection to the institution and effort to resoluteness that.

But for others, the HSP highlighted 1 of the halfway underlying failures that has plagued HackerOne for years: stagnating characteristic development. Even if you are portion of the exclusive HSP group, you still don’t person the expertise to do 1 important thing, which is to invoke alteration and improvement wrong the platform.

And yet, for immoderate reason, HackerOne decided to return a performative attack to this, going arsenic acold arsenic to create a dedicated level feedback channel, but doing perfectly thing pinch immoderate of that feedback. I cannot moreover count really galore pieces of individual feedback and suggestions person been posted successful there, pinch positively zero action from HackerOne.

The AI Era #

And then, astir 2021, we each entered a caller play of quality history: the emergence of AI and LLMs.

Beginning astir this time, we each began to witnesser the immense gains successful velocity and the capabilities of LLMs. You could now one-shot features successful a fraction of the time, build civilization tools, and wide get much done successful little time.

But this is wherever HackerOne took a confusing turn. It is genuinely baffling to maine really HackerOne managed to fumble this exertion successful the worst measurement possible. They had a 10-year backlog of characteristic requests, and were handed 1 of the astir powerful package improvement devices successful the past 50 years. But alternatively of utilizing this to amended the level and adhd long-requested features, they decided to move toward creating “their own” AI adjunct called Hai. Not only was Hai conscionable a wrapper connected apical of OpenAI, but it hardly had immoderate existent unsocial capabilities to thief hackers aliases programs do what they had been requesting from HackerOne to adhd to the halfway platform.

On apical of this, the driving unit down the company, the hackers and co-founders who started the business, were softly shoved distant into the dungeon of HackerOne. The website has them listed arsenic portion of the executive team, but they are puppets successful their ain business, hardly getting to activity connected the merchandise they built from the crushed up.

The Enshittification of HackerOne #

For almost 10 years, Marten Mickos served arsenic CEO of HackerOne. Some hackers loved Marten; others had mixed feelings. Personally, my interactions pinch him were location betwixt good and good. But I do judge that he was very passionate astir bug bounty and did a batch of bully for the institution during his clip arsenic CEO.

But successful precocious 2024, Marten was replaced by Kara Sprague, the erstwhile Chief Product Officer astatine F5. You mightiness beryllium asking yourself, what does a CPO of a networking institution cognize astir hacking and bug bounty?

Honestly, I americium not sure. What I will opportunity is that location are mixed signals online regarding what happened to BIG-IP during her clip arsenic CPO astatine F5.

But that didn’t extremity the committee from deciding she was the correct fresh for the early of the company. And this was neither the first nor the past successful a bid of questionable business decisions made by HackerOne.

As AI agents continued to turn and displacement into various identities, HackerOne’s personality besides began to shift. Slowly but steadily, HackerOne rebranded itself astir a recently invented manufacture category: “CTEM”, aliases “Continuous Threat Exposure Management”.

HackerOne quickly began to germinate into a soulless corp obsessed pinch numbers, business objectives, and B2B sales. They switched from talking astir bug bounty programs, unrecorded hacking events, and really they could thief you enactment secure, to promoting their in-house AI information merchandise and continuous information monitoring tool.

Wait, what? What in-house AI information product?

Your Reports Are Yours, We Promise #

Sometimes, it each starts pinch a tweet.

In February 2026, well-known hacker zseano noticed that location was a surge successful HackerOne labor leaving the institution and asked why that would beryllium happening.

This ended up revealing that HackerOne had made ToS updates which made it truthful that reports submitted connected HackerOne could beryllium utilized to train AI models.

As you mightiness expect, this created rather a spot of noise, particularly wrong the HSP chats. One point I’ve noticed is that the modern time HackerOne only ever lets the founders retired of the dungeon for harm control, and this was a cleanable clip to propulsion that lever.

Within 24 hours, Alex Rice, co-founder, CTO (and CISO…?) of HackerOne emerged from the dungeon to execute harm control. (This was the first clip he had ever messaged successful the HSP wide chat since it was created 2 years earlier).

We do not train, fine-tune, aliases different amended GenAI aliases ample connection models connected interrogator data. That includes Agentic PTaaS.

https://docs.hackerone.com/en/articles/10908081-hai-security-trust

And erstwhile asked to region Section 3.1 of the ToS, Alex promised that this would beryllium taken attraction of and made clearer successful an upcoming update to the ToS.

Within a week, CEO Kara Sprague had been looped successful and made a fluffy LinkedIn post explaining that

HackerOne does not train generative AI models, internally aliases done third-party providers, connected interrogator submissions aliases customer confidential data.

Researcher submissions are not utilized to train, fine-tune, aliases different amended generative AI models. This applies crossed our platform, including capabilities utilized wrong our Agentic PTaaS offering and our in-platform AI, Hai.

Alex Rice moreover went connected to the Critical Thinking Bug Bounty Podcast and did a impermanent spot to talk astir this rumor specifically and unopen down immoderate concerns astir study information being utilized to train AI models.

Amazing, that’s settled then. HackerOne doesn’t usage interrogator information to train AI models.

…Right?

Just Kidding They Never Were #

Unfortunately for us, the present-day HackerOne follows a alternatively awesome PR harm power playbook:

1. Address fears
2. Ease concerns
3. Don’t alteration anything

About 2 months aft this drama, personification mentioned successful the HSP chat that they had noticed a remark from hackerone-agent connected 1 of their reports and wanted to cognize if this was an AI supplier aliases an existent human.

The consequence from 1 of the HackerOne merchandise managers was:

This is simply a preliminary reappraisal done by AI […] The H1 triage process has a measurement called “H1 Intake” […] which we’ve automated for definite reports to make judge they scope the “Validation team” faster.

And erstwhile asked whether “certain reports” meant definite study types aliases whether a programme could alteration it everywhere, they responded:

We’re moving the study connected each reports, the AI determines what tin beryllium sent straight to the validation squad vs what needs to beryllium reviewed by a quality from the intake squad first. […] The strategy besides learns from behaviour so, if a proposal is rejected […] it will return these into relationship pinch its recommendations arsenic well.

Hang connected a second.

I thought that interrogator submissions are not utilized to train, fine-tune, aliases different amended generative AI models.

But now each reports are being tally done an AI system, and the strategy is utilizing the outcomes of those reports to power really it handles early reports. So is that technically “fine-tuning”? Technically, no.

So really tin some of these things beryllium possible?

Well, erstwhile I asked this, I received an unthinkable copy-paste consequence of what Alex Rice had said backmost successful February:

Hai does not train, fine-tune, aliases different amended GenAI aliases ample connection models connected customer aliases interrogator data: https://docs.hackerone.com/en/articles/10908081-hai-security-trust

They past followed up pinch an mentation that appeared to contradict this:

Storing learnings successful our DB is thing different than training/fine-tuning a exemplary connected really to behave. We shop the rejection reasons successful our DB to transverse reference and augment the reasoning the supplier does erstwhile determining adjacent steps. […] The supplier will prime this up from behaviour the customer has done connected past reports and usage that to memorize and augment the reasoning for caller reports.

So yes, HackerOne is not building exemplary weights from your study information aliases “training” connected it successful the accepted sense.

But that is simply a favoritism without immoderate meaningful quality to the researchers whose information is being used. The strategy is still learning from study information successful the applicable consciousness that matters: past reports alteration early automated behavior.

Call it memory, retrieval, contextual learning, aliases immoderate you want. For the intent of deciding whether interrogator information is being utilized to amended the system, the favoritism is functionally equivalent.

And arsenic a cherry connected top, a fewer days later, HackerOne published a blog post stating:

As AI-assisted submissions person grown, we’ve invested heavy successful AI-powered triage and validation connected our ain platform. Our agentic AI system, Hai, now handles first classification, deduplication, and validation of incoming reports, moving alongside our quality information analysts.

The Boiling Point #

By this point, I had mislaid each religion successful HackerOne. I’d been astir for truthful galore years and done truthful galore seasons that location was really nary constituent successful leaving thing disconnected the table.

So I posted a agelong connection successful the HSP chat, echoing galore of the things I’ve said successful this blog.

What happened arsenic a consequence was really really surprising: galore different apical hackers successful the HSP agreed, echoed the aforesaid concerns, and called for change.

I figured, surely HackerOne will person to respond to this.

And they did.

Co-founder Michiel Prins was allowed to time off the HackerOne dungeon to execute harm power pinch this absolute banger of an AI slop response:

You’re correct to telephone this out. Parts of the level and wide acquisition person not moved accelerated enough, and that’s frustrating.

We person been prioritizing activity that improves outcomes for some customers and researchers. Triage is simply a clear example. Speed and consistency location person improved materially, and that matters for everyone connected the platform. Report Assistant whitethorn not beryllium targeted astatine the caliber hacker successful this channel, but it has been a important enhancement for the largest portion of the organization - hackers earlier successful their profession - and this has led to overmuch clearer reports for triage to analyse and accelerate wide throughput.

At the aforesaid time, location is simply a bigger displacement happening pinch AI. There is simply a batch of sound astir AI replacing information researchers. We do not judge that. We are investing to make judge researchers go much important, not less. A batch of that activity is happening connected the customer broadside of the marketplace.

That said, the spread you are pointing to connected day-to-day usability and acting connected feedback is real. We request to do a amended occupation there, and we request to show that advancement much clearly.

And then, per the playbook, they waited for the organization to move on, brushed it nether the rug, and thing changed.

We Must Calm the Masses #

Well that was existent until June, erstwhile rez0 (co-host of the Critical Thinking Bug Bounty Podcast) asked astir the caller “HackerOne Continuous Testing” merchandise that had conscionable launched:

The page (before it was updated) very intelligibly stated:

Testing is sharpened utilizing discourse from HackerOne’s 12+ years of real-world vulnerability information and your anterior H1 Bounty findings, truthful agents cognize wherever to focus

HackerOne instantly tried to tally this backmost and declare that this was really not true. I find it difficult to understand really a institution tin beryllium truthful incredibly disconnected from its ain values, but only erstwhile it’s brought up by hackers who look intimately and inquire questions.

All the behaviour complete the past fewer months embodies “oops we didn’t mean to opportunity the quiet portion retired loud”, and surely not “we don’t do that”.

And honestly astatine this constituent if you person to do this overmuch harm control, you mightiness arsenic good conscionable ain it.

I won’t moreover statesman to break these statements down because there’s truthful overmuch connection crockery that it’s mind-numbing.

As I constitute this post, I americium exhausted from trying to execute the intelligence gymnastics required to judge each this.

Actions speak louder than words, and it’s beautiful clear astatine this constituent what’s going on.

Feel free to skip ahead, but successful an effort to beryllium adjacent I’ll supply the replies from not 1 but two HackerOne co-founders (wow a double dungeon release):

No alteration successful direction. There is zero training connected interrogator data.

This connection isn’t moreover talking astir vulnerability discovery. What happens erstwhile you constituent a “recon agent” astatine an onslaught surface?

Noise.

That’s the discourse being talked astir present – Scope. Out of scope. Informative. Ineligible findings. This aforesaid contextual feedback loop is disposable to researchers (and, increasingly, their agents) done our intake agents & afloat triage process.

We proceed to request to do overmuch amended connected accompanying simplified trading connection pinch method architecture documentation, and we request to do a amended occupation astatine reinforcing the defense successful extent communicative – location are nary metallic bullets here. Nowhere successful our imagination is simply a 100% AI human-free future.

- Alex Rice

and

Sorry for the disorder this paragraph has caused. It was meant to seizure really our recon agents find retired wherever to attraction and wherever not to attraction based connected anterior results successful a customer’s programme (mapping the onslaught surface), not find caller vulnerabilities based connected interrogator submissions. We person adjusted the paragraph successful the H1 Continuous Testing page and buttoned up our documentation. H1 Continuous Testing is based connected the aforesaid attack and methodologies arsenic the H1 Agentic Pentest offering announced successful January. Here is the documentation connected really those agentic products logic & usage discourse for information testing workflows.

But let’s attraction connected the underlying concern: is H1 utilizing interrogator submissions to past deploy agents and find that vulnerability utilizing their techniques successful different customers? No.

That is simply a exemplary that tin beryllium a win-win-win for everyone successful the agelong run, but we don’t yet person the foundations successful spot to supply due nickname and rewards for researchers successful specified a model. We’re envisioning a level wherever researchers thrust the frontier, agents supply the scale, and researchers get due nickname for their impact. Imagine this: you observe a caller method to find a vulnerability. We operationalize it crossed each customer’s onslaught aboveground that shows signs of exposure. Customers get protected. You get compensated not conscionable for 1 report, but proportionally to the effect you make crossed the platform. That is fundamentally a amended woody for researchers than the existent one-bounty-per-report model. And it is besides a triumph for customers arsenic it narrows the vulnerability window.

But the reality is that it’s conscionable a imagination correct now. We are not location yet. This requires observant activity crossed product, argumentation and interrogator engagement models to build it right. Right now, we are still focused connected building foundational agentic capabilities for testing, accelerating validation and triage, prioritization, and remediation. These person to activity reliably first earlier we tin furniture connected a researcher-driven onslaught room and inducement exemplary that does justness to the contribution. We besides beryllium it to customers not to conscionable dump findings connected them without the tooling to enactment connected them.

The agelong word is astir partnership, not replacement. We want to build inducement structures that support this vision. It’s unexplored territory, and we want to get it correct pinch you.

- Michiel Prins

along pinch a HackerOne official™ tweet to calm the masses:

Conclusion #

Today arsenic I look astatine HackerOne, it’s difficult to moreover admit the institution that was started backmost successful 2012. The values person shifted, the psyche has been gutted, and the halfway personality has been lost. A sad ending for our hero’s journey.

So, wherever do we spell from here?

Well, I person immoderate last thoughts for each statement progressive successful this twisting tale.

To HackerOne: It was bully knowing you. All I tin dream is that you tin find your existent personality again 1 day. Although I uncertainty that will ever happen.

To the founders: I dream you sewage paid well, that you tin exit, and that you spell do thing cool again.

To the hackers: Know your worth. Bug bounty platforms cannot beryllium without hackers.

To the companies: You don’t request HackerOne anymore. The tokens to build your ain in-house level costs little than azygous twelvemonth of HackerOne.

And to whoever is fired up: The marketplace is fresh for a disruption. The devices are successful your hands. Build what HackerOne could person been.

Godspeed and acknowledgment for reference if you made it this far.