---
type: source-pointer
speaker: Valentina Palmiotti (chompie)
affiliation: IBM X-Force
title: "BinChomp — autonomous n-day exploit generation pipeline"
event: "[un]prompted.au 2026 (Sydney)"
event_date: 2026-09-18
event_role: featured talk
event_url: https://www.unprompted.au/schedule
event_archive_url: null
video_url: null
video_archive_url: null
video_url_status: "No recording verified during this pass."
slides_url: https://www.beautiful.ai/player/-P2q0r-B7nGYGGK3-32Z/Binchomp-N-Day-Pipeline-Pres
slides_archive_url: null
archive_note: "No matching third-party archive snapshot was verified; null is not evidence that none exists."
date_caveat: "The official schedule places the matching title and speaker on Day 1, 18 September 2026, at 13:30 AEST. The supplied 36-slide player was inspected on 17 October 2026; its exact revision date and identity with the version presented at that event are unverified."
identification_confidence: high
identification_evidence: |
  The official event schedule identifies Valentina Palmiotti, IBM X-Force, and
  BinChomp. The user-supplied player identifies the Windows-kernel N-day pipeline
  and was read through its closing slide; economic tables and limitations were
  checked visually because the player accessibility text omits embedded figures.
integration: |
  Used in research/2026-10-17-discovery-costs/RESEARCH.md and
  2026-discovery-costs.html. Structured observations are in
  data/discovery-cost-evidence-2026-10.json, row DC01. These research observations
  do not calibrate the interactive pricing model or change its multipliers.
related_corpus_rows: []
relevance: |
  A first-party case study of the marginal API expense of developing outputs for
  known Windows-kernel vulnerabilities. It separates reproduction from verified
  chains and highlights the contribution of reusable tooling and researcher
  judgment. It supplies no transaction price or all-in discovery cost.
license: "Valentina Palmiotti / the relevant rights holders retain rights to the presentation; this file contains metadata and nonverbatim integration notes, not a transcript."
last_verified: 2026-10-17
---

# Valentina Palmiotti — BinChomp

The [official schedule](https://www.unprompted.au/schedule) identifies the talk at [un]prompted.au on 18 September 2026. The [supplied slides](https://www.beautiful.ai/player/-P2q0r-B7nGYGGK3-32Z/Binchomp-N-Day-Pipeline-Pres) are a live player, so the numbers below describe the version inspected on 17 October, not a frozen conference artifact.

## Evidence used

- **Slides 13 and 30 — task and sample.** The pipeline starts from known CVEs. Its evaluation uses 21 Windows-kernel CVEs from Microsoft's January and February 2026 updates. The presenter reports 17 successful reproductions and six verified privilege-escalation chains with Opus 4.7. These are different completion criteria, not 23 independent discoveries.
- **Slide 31 — spending boundary.** The reported $1,098 is the experiment's API bill, including reproduction, unsuccessful CVEs and all chain-building attempts. Dividing that bill by six verified chains gives the displayed $183 per chain. The deck does not price setup, researcher time, reusable engineering, tools, infrastructure or monitoring.
- **Slides 30–31 — comparator boundary.** The deck quotes [Anthropic's 8 June 2026 study](https://www.anthropic.com/research/n-days): Opus 4.7 reproduced 13/21 with no full chains; Mythos Preview reproduced 18/21 and completed 8/21 chains. Anthropic's Windows results independently confirm these counts and its separately reported API totals of $15,700 for chain work and approximately $2,200 for reproduction. The deck flags uncertainty about inclusion of failed-CVE spending; the original does not resolve that accounting detail. BinChomp used a CVP-verified organization account with guardrails enabled; the comparator used disabled safeguards and three trials per CVE. These differences prevent a clean efficiency ranking.
- **Slides 25, 28 and 32 — reusable expertise.** Existing libraries and an advisory agent informed by the researcher's earlier work structure the task. Development reruns suggest fewer retries and failures, with shorter development when libraries are reused. The slide explicitly describes qualitative observations; it provides no effect size, numerical elapsed time or controlled sample count.
- **Slides 33–34 — access and supervision.** The author identifies model access as an evaluation constraint. The bespoke prototype was manually monitored throughout. Its outputs were minimum viable demonstrations, with no focus on stealth or defensive-tool evasion, and may not work in heavily guarded environments.

## Integration boundary

The case supports a conceptual shift from repeated manual execution toward reusable workflows, verification and accumulated expertise. That is an interpretation of a small first-party experiment. It does not establish a general time trend, an all-in cost per new vulnerability, a market clearing price, or the price of a durable operational capability. A verified local demonstration should remain distinct from a reliable, maintained product and from access in a live environment.

See the [research synthesis](../research/2026-10-17-discovery-costs/RESEARCH.md), [discovery-cost edition](../2026-discovery-costs.html) and [typed evidence](../data/discovery-cost-evidence-2026-10.json).

## Open provenance gaps

The exact deck revision date, number of BinChomp trials per CVE, per-CVE spending distribution, wall-clock measurements and monitoring hours are not established by these slides. Archive URLs remain null pending verification; no archive or recording is invented.