November 12, 2026
The 610-Character Report Paid $8,000.
The crash class was the same in December and March. The price wasnโt.

By Raj Namdev
6 min read
The 610-Character Report Paid $8,000. The 5,419-Character One Paid $800. The Difference Was One Rating.
30-second version:_ One HackerOne program โ Shopify's mruby sandbox โ has 161 public reports, 88 of them paid, $251,300 in total. Two-thirds of that money sits on 22 reports disclosed in a six-day window in December 2016, when the median paid report was $8,000. Four months later, a report with the same symptom โ a null-pointer crash in class handling โ and nine times the write-up was rated low and paid $800. The payout schedule had nine distinct amounts; the severity label, not the prose, picked the rung. The $18,000 top rung went to a 710-character report that proved instruction-pointer control._
The first paid report in this program was disclosed on December 16, 2016. By December 21, the reports disclosed in that window carried $167,000 in bounties โ 66.5% of every dollar in the program's public record.
The median paid report in those six days: $8,000. Across the program's whole life: $800.
I pulled all 161 reports as JSON and read them. Not the summaries โ the bodies, the severity ratings, the bounty amounts, the submission dates. The interesting thing wasn't the money. It was how mechanically the money followed a label.
What this program is, and how I pulled it
shopify-scripts is Shopify's public program page for the mruby interpreter and its sandbox wrapper, mruby-engine. The reports are interpreter bugs โ crashes, memory corruption, type confusion โ filed between December 2016 and September 2019.
Both files are public: the reddelexc/hackerone-reports corpus for the report list, HackerOne's JSON for everything else.
for id in 183405 212107 215891 215447; do
curl -s "https://hackerone.com/reports/$id.json" \
| jq -r '[.id,.severity_rating,.bounty_amount,.disclosed_at,.title]|@tsv'
donefor id in 183405 212107 215891 215447; do
curl -s "https://hackerone.com/reports/$id.json" \
| jq -r '[.id,.severity_rating,.bounty_amount,.disclosed_at,.title]|@tsv'
doneThe median paid report has 4 upvotes. The audience was tiny, which is part of why the record is clean.
Nine amounts, six of them covering 84 reports
The payouts aren't a curve. They're a price list.
$100โ16 reports (18% of paid) $200โ2 reports (2%) $300โ1 report (1%) $800โ35 reports (40%) $1,000โ11 reports (13%) $2,000โ1 report (1%) $8,000โ10 reports (11%) $10,000โ11 reports (13%) $18,000โ1 report (1%)
Six amounts cover 84 of the 88 payments. The other four landed on $200, $300 and $2,000. There is no $3,400, no $6,500, no negotiated middle โ prices that cluster in steps, not in a curve, which is what a rubric looks like when it leaks into public data.
The six days
Group the paid reports by disclosure month and the shape gets obvious.
2016โ12โ22 paid, $167,000 total, median $8,000 2017โ01โ8 paid, $21,800 total, median $900 2017โ02โ17 paid, $8,700 total, median $200 2017โ03โ9 paid, $23,700 total, median $800 2017โ04โ14 paid, $17,200 total, median $800 2017โ05โ7 paid, $4,200 total, median $800 2017โ06โ2 paid, $900 total, median $450 2017โ07โ1 paid, $800 total, median $800 2017โ12 to 2019โ09โ8 paid, $7,000 total, median $800
Two honesty notes before this becomes a story about a program getting cheap. Disclosure date is not payment date โ a report can be paid at triage weeks before it goes public. The December cohort was submitted between November 10 and December 10, so the disclosures cluster because reporters agreed to publish together. What the window measures is the cohort: the first wave of reports, priced while the program was new. I can't see Shopify's policy history. I can see what they paid, to whom, and what label they attached.
The $8,000 crash and the $800 crash
Here's the pair that made me go through the rest.
#183405, "Null target_class DoS," submitted November 19, 2016, disclosed December 17. Rated high, paid $8,000, body length 610 characters. The whole technical argument:
If a singleton class cannot be created for an object,
target_classis set to NULL. TheOP_CLASSandOP_MODULEopcodes in the VM assumetarget_classis not null when defining new classes and modules. This causes a null pointer dereference and segfaults the mruby VM.
1.instance_exec { class X; end }
#212107, "Null pointer dereference in mrb_class," submitted March 10, 2017, disclosed March 14. Rated low, paid $800, body length 5,419 characters. Same interpreter, same symptom, nine times the write-up.
Read side by side, nothing in those bodies explains a 10x gap. The ratings differ, and the ratings set the price.
The sequence continues. #215891, another null-pointer crash in mrb_class, April 2017: $800, rated low. #215447, "SIGSEGV in mrb_class," submitted March 2017: $0, no rating.
Four reports, five months, class handling: $8,000, $800, $800, $0. I can't prove from the public JSON that they share a root cause โ only one duplicate relationship in the whole program is marked โ but that's what saturating a bug class looks like from the outside.
# counts and the exact amounts paid, per severity label
jq -s 'map(select(.bounty_amount>0))
| group_by(.severity_rating)
| map({sev:.[0].severity_rating, n:length,
amounts:(map(.bounty_amount)|unique)})' rep/*.json# counts and the exact amounts paid, per severity label
jq -s 'map(select(.bounty_amount>0))
| group_by(.severity_rating)
| map({sev:.[0].severity_rating, n:length,
amounts:(map(.bounty_amount)|unique)})' rep/*.jsonThe label-to-price mapping is nearly total. high pays $8,000 or $10,000 (21 of 23). critical is the single $18,000. low pays $800 or $1,000 (17 of 21). Unrated reports sit at $800 (21) or $100 (5). Across all six labels, 77 of the 88 payments land in the two amounts their label pays most often. The other 11 scatter between $100 and $2,000.
What the $18,000 report actually proved
The top of the ladder is one report, and it's 710 characters.
#181879, "Struct type confusion RCE," critical, $18,000, disclosed December 17, 2016 โ two days after the program's first public report. Its opening line:
I've been poking at mruby a bit more and I've found a vulnerability that allows an attacker to take control of the instruction pointer.
The PoC jumps to 0x0000133713371337 and crashes. That's it. Fourteen upvotes, a 710-character body, and an attached script that moves the instruction pointer to an address the author chose.
Compare that to #185957, "Crash: A call to Symbol.new leads to a crash when inspecting the resulting object," disclosed the day before: 39,416 characters, rated low, paid $1,000.
One day apart, thirty-nine thousand characters apart, and the short one made eighteen times as much. Length wasn't the ladder. Control was.
Across the whole program, 33 paid reports have bodies of 1,000 characters or less, and they hold $78,600**โ31.3%** of every dollar paid. Brevity wasn't punished.
What I'd check before writing the next crash report
The prices here are stale; the shape is transferable. In this record, a payout tracked two things: the primitive you could prove, and whether the root cause was new. Both are visible before you write a word.
- Name the primitive, not the symptom. "Segfault in the VM" is a DoS. "Control of the instruction pointer" is not. The $18,000 report is that second sentence plus a PoC; the $800 ones document the crash.
- Search the program's public history for your class first. Class-pointer crashes paid $8,000, then $800, $800 and $0. Checking the first three would have priced the fourth before the work, not after.
- Read the tier table before you argue about severity. "This should be critical" is an opinion. "This is control of a pointer, which your program has paid at the $8,000 tier for the same primitive" is a comparison.
The uncomfortable part: the December cohort wasn't paid for prose. They filed in the right month, against the right code, with the right primitive.
What this doesn't prove
The 66.5% figure is a disclosure-window artifact as much as a pricing one. Reports get disclosed when reporters agree to publish, and the first cohort of a new program tends to publish together. Payment dates aren't in the JSON, so I'm describing when reports became public, not when money moved.
Severity ratings aren't documentation either. HackerOne added structured ratings over time, and older reports were rated retroactively or not at all โ 31 of the 88 paid reports carry no rating, and those have a median of $800.
The title matching is loose. I grouped reports by the words in their titles โ "null pointer," "segfault," "crash" โ a symptom match, not a root-cause match. Only five reports are marked duplicates, and those paid $0, so the record can't tell you how many of the $800 crashes repeated an earlier root cause.
Fourteen paid reports have no public body โ the reporter chose no-content visibility. Everything I say about write-ups comes from the 74 paid reports with a readable body.
The last line of the table
The program's most-upvoted report has 60 upvotes. The $18,000 one has 14. The $800 tier has a median of 3 upvotes.
Sort this program by attention and you learn almost nothing about where the money was. Sort it by severity label and 77 of 88 payments land within two rungs.
The crash class was the same in December and March. The price wasn't.
Sources
Pulled November 12, 2026. Every figure was computed locally from the raw JSON; no number is reused from an earlier draft.
- Corpus โ
raw.githubusercontent.com/reddelexc/hackerone-reports/master/data.csv, filtered toprogram == "shopify-scripts": 161 rows. - Report JSON โ
hackerone.com/reports/<id>.json, all 161 IDs, 161/161 resolved. - Program totals โ 161 reports, 88 paid, $251,300, median $800, max $18,000. Histogram: $100ร16, $200ร2, $300ร1, $800ร35, $1,000ร11, $2,000ร1, $8,000ร10, $10,000ร11, $18,000ร1.
- December window โ 22 paid reports, disclosed 2016โ12โ16 to 2016โ12โ21, $167,000 (66.5%); submitted 2016โ11โ10 to 2016โ12โ10.
- #183405 โ "Null target_class DoS," high, $8,000, 610-char body. #212107 โ "Null pointer dereference in mrb_class," low, $800, 5,419-char body. #215891 โ same title, $800, 2017โ04โ15. #215447 โ "SIGSEGV in mrb_class," $0.
- #181879 โ "Struct type confusion RCE," critical, $18,000, 710-char body, 14 upvotes. #185957 โ Symbol.new crash, low, $1,000, 39,416-char body.
- Body split โ 33 paid reports โค1,000 chars, $78,600 (31.3%); 14 paid with no-content; 31 paid with
severity_rating: null.