Static leads and a verified RCE

Granite’s recorded investigations produced static candidate inventories and code-level evidence. Our team subsequently verified one RCE dynamically using a required test harness. These are distinct outcomes. [1] [2]

Dynamically verified by our team

1RCE dynamically verified

Required test harness

Our team dynamically verified one remote code execution finding under the harness setup. Other findings remain static candidates. [2]

Recorded separately from the static inventories, without target or workflow attribution.

Read the verification scope

Static candidates · Target A

19static research leads

Granite Desktop’s IDs form 18 provisional root-cause families and include conditional items. They are not a verified-vulnerability count. [1]

Examine Target A’s findings

Static candidates · Target B

Static candidates across both Granite workflows

Granite Desktop
16
Granite MCP
13

Reported static out-of-bounds candidates from separate workflows. The inventories are not added together. [1]

Examine Target B’s findings

These are seven workflow arms across two datasets, not seven independent target trials. Target names, versions and identifying artifacts are withheld. Each dataset stands on its own; function receipt, static candidates and dynamic validation measure different things.

Target A

Three recorded workflows

Three exploratory sessions with different durations, instructions and worker resources. Direct code/assembly and additional description-only functions are shown separately. The counts describe material returned to the agents, not equivalent completed investigations. [1]

Reported static candidates

Retained static IDs · conditional items included

  1. Granite Desktop19
  2. REAopencode3
  3. IDA MCPopencode1

Finding criteria differ across workflows. These counts do not form a quality ranking or a count of dynamically verified vulnerabilities.

Investigation outcomes · Target A

Follow-up that ruled out a hypothesis

Candidate inventory
19 Granite static IDs form 18 provisional root-cause families.
Hypothesis refuted
A follow-up branch refuted one hypothesis.
Question unresolved
Another question remained unresolved after human guidance.

Recorded research outcomes; analyst review time was not measured. [1]

Direct code or assembly

Returned function counts

  1. Granite Desktop657
  2. REAopencode24
  3. IDA MCPopencode43

Recorded model usage

Raw agents + engine, USD

  1. Granite Desktop$14.16
  2. REAopencode$0.81
  3. IDA MCPopencode$1.45

Corrected rates. Setup, licensing, infrastructure and analyst review are outside these totals.

Returned evidence and static candidates

Scroll to compare columns
Target A — returned evidence and static candidates
WorkflowDirect functionsDescription-onlyCombined unionStatic IDs
Granite Desktop6577351,39219
REAopencode240243
IDA MCPopencode430431

The 19 Granite static IDs form 18 provisional root-cause families and include conditional items. The other workflows report three conditional findings and one principal static finding; one location overlaps a Granite finding. These counts use different finding criteria and do not form a quality ranking. [1]

Recorded resources and corrected model-use equivalents

Scroll to compare columns
Target A — recorded resources and corrected model-use equivalents
WorkflowElapsed minutesCode-bearing surfacesAgentsUSDEngineUSDTotalUSD
Granite Desktop156.921$12.86$1.30$14.16
REAopencode55.96$0.81$0.00$0.81
IDA MCPopencode30.46$1.45$0.00$1.45

Granite’s corrected total is $14.16: $12.86 agents plus $1.30 engine. Title generation, cancelled usage and unsuccessful research remain included. A follow-up branch refuted one hypothesis and left another question unresolved after human guidance. Analyst review time was not measured. [1]

Granite used 15 descendant contexts alongside its parent context; peak simultaneous workers are not reported. IDA MCP had a one-worker ceiling. Code-bearing surfaces include support and local-path code, not just entry points or independent products.

Target B

Four recorded workflows

Four recorded workflows, with complete bodies, excerpts, assembly and semantic descriptions counted as separate evidence tiers. The earlier two- and three-workflow reports reuse these same arms; they are included here once. [1]

The REA-labelled workflow is a hybrid: it temporarily used Granite pseudocode extraction. Script-only scans, names and sink flags do not count as returned function bodies. The workflows disagree on some candidates; these counts retain each workflow’s classifications. [1]

Reported static OOB candidates

Out-of-bounds candidates · not dynamic confirmations

  1. Granite Desktop16
  2. REA-labelledHybridopencode2
  3. Granite MCPopencode13
  4. IDA MCPopencode5

Each bar retains its workflow’s reported classifications. Candidate verdicts conflict across workflows and have not been adjudicated.

Investigation outcomes · Target B

16 Desktop candidates. 13 through MCP.

Granite Desktop
16 reported static out-of-bounds candidates.
Granite MCP
13 reported static out-of-bounds candidates.
Cross-workflow review
The workflows disagree on some candidates. The counts retain each workflow’s classifications.

The public data does not assign the separately verified RCE to a target or workflow row. [1] [2]

Complete function bodies

Returned function counts

  1. Granite Desktop755
  2. REA-labelledHybridopencode26
  3. Granite MCPopencode337
  4. IDA MCPopencode600

Recorded model usage

Raw agents + engine, USD

  1. Granite Desktop$5.56
  2. REA-labelledHybridopencode$3.82
  3. Granite MCPopencode$5.20
  4. IDA MCPopencode$9.63

Raw usage shown. IDA MCP’s analytically adjusted total is $7.32; the deduction is detailed below. Setup and analyst review are outside these totals.

Investigation depth and reported static candidates

Scroll to compare columns
Target B — investigation depth and reported static candidates
WorkflowComplete bodiesAny code / assemblyBodies or descriptionsStatic OOB candidates
Granite Desktop75576693816
REA-labelledHybridopencode2682262
Granite MCPopencode33733966213
IDA MCPopencode6007236005

Raw and analytically adjusted model-use equivalents, USD

Scroll to compare columns
Target B — raw and analytically adjusted model-use equivalents, USD
WorkflowAgentsEngineRaw totalAdjusted total
Granite Desktop$4.84$0.72$5.56$5.56
REA-labelledHybridopencode$3.82$0.00$3.82$3.82
Granite MCPopencode$4.56$0.63$5.20$5.20
IDA MCPopencode$9.63$0.00$9.63$7.32

For IDA MCP, the adjusted view deducts $2.31277386 for 209 model turns whose tool calls returned only a capacity error. A mixed turn with a successful write remains charged. The raw total is $9.62812536; the adjusted total is $7.31535150. This is an analytical deduction from the same run, not a rerun or measured savings from a different configuration. Other rows are unchanged. [1]

Detailed counting tiers and function sizes

Additional evidence tiers

Scroll to compare columns
Target B — additional evidence tiers
WorkflowExcerpt-onlyAssembly-onlyDescribedincl. overlapAny code or descriptions
Granite Desktop47555949
REA-labelledHybridopencode515082
Granite MCPopencode11555663
IDA MCPopencode12030723

Complete-body content and surface counts

Scroll to compare columns
Target B — complete-body content and surface counts
WorkflowBody charactersMedian charactersBodies under 200 charactersSurfaces with bodies
Granite Desktop2,921,4311,9869139
REA-labelledHybridopencode82,7371,366010
Granite MCPopencode1,542,3983,0352039
IDA MCPopencode1,609,2921,5584139

Union columns deduplicate functions across their stated tiers. A description is contextual information, not another complete body. Body length and short-function counts provide sensitivity views, not correctness scores. The data download preserves exact numeric values. [1]

One RCE verified. Harness required.

Verification scope

Our team dynamically verified one RCE using a required test harness. Verification followed the recorded static benchmark sessions and applies to that harness setup. [2]

The tables retain their static candidate counts; the dynamically verified result is recorded separately.

Target names and issue-specific reproduction details are excluded from this public benchmark dataset.

The single verified result cannot establish a confirmation rate for the candidate inventory, recall or a false-positive rate. Target identities remain withheld for both datasets.

Run conditions and accounting

All tables use corrected supplied token rates. Agent uncached input / cached input / output rates are $0.50 / $0.08 / $2.00 per million tokens; engine rates are $0.10 / $0.01 / $0.50. Values are usage equivalents, not verified invoices, customer prices or full task costs. Infrastructure, subscriptions, licensing, setup and analyst review are outside these denominators. [1]

Cost per function is an auxiliary process measure. A function receipt does not establish complete review, and different workflows pursued different depths of investigation. Function counts, candidate counts and descriptions cannot be treated as interchangeable units of useful work.

Prompts, stopping rules, worker resources, setup delays and target coverage differ. The recordings do not support a controlled causal claim, pooled detection rate or confidence interval based on individual functions. A controlled comparison requires matched target-task pairs, each system’s intended orchestration, and blinded assessment of accepted outcomes and reviewer effort.

Accounting audits check the packaged usage and credited ledgers, not vulnerability validity or comprehension. Target B’s source accounting audit passes 59 checks. Retained session exports omit streaming deltas and repeated request fields; the public download contains aggregate measurements, not original traces. [1]

Data & sources

This page consolidates both current benchmark datasets and all seven recorded workflow arms from our benchmark repository. Superseded pricing and provisional attribution are replaced by the current accounting; adjusted views are shown alongside raw values. Individual target identifiers, binary names and finding artifacts are withheld.

  1. Anonymous benchmark data, measurement snapshot 17 November 2026. Includes both datasets, every workflow row, counting tiers, costs, adjustment and rate assumptions. Source references A and B correspond to retained internal reports and accounting audits.
  2. Binary Scryer validation update, 18 November 2026: our team dynamically verified one RCE using a required test harness. The public dataset records this outcome and its harness requirement; issue-specific reproduction records are not included.

Glossary

Static analysis
Examining code without executing the behavior under study.
Dynamic verification
Checking a claimed behavior by executing it under stated conditions.
Test harness
Supporting code or setup used to exercise and observe behavior under test.
RCE
Remote code execution. The verified result requires a test harness.
OOB
Out-of-bounds memory access. A static OOB candidate is not automatically a dynamically verified vulnerability.
Root-cause family
Reports grouped because they describe the same underlying defect or mechanism.
Refutation
Evidence contradicting a proposed explanation.
Observational comparison
Recorded runs compared without controlling all relevant conditions.
Recall and false-positive rate
Measures requiring independently established positive and negative reference cases, unavailable in these datasets.
Download the dataExplore Granite