The benchmark sampled known Python web-application CVEs, confirmed original exploits and then tested older releases that vulnerability databases did not identify as affected. Agents were asked to migrate PoCs when the old source, API or dependency structure made a direct replay ineffective. Researchers verified 149 older affected historical releases across…
Inspect the ClaimsCase · DiggingBeagle record
CrossPloit used executable tests to show 149 older vulnerable releases omitted from advisory version ranges
The benchmark sampled known Python web-application CVEs, confirmed original exploits and then tested older releases that vulnerability databases did not identify as affected. Agents were asked to migrate PoCs when the old source, API or dependency structure made a direct replay ineffective.
Public disclosure: Nov 12, 2026
An agent has access to a known exploit and older software versions whose API/implementation drift affects applicability.
At a glance
An execution-checked migrated PoC extends the effective vulnerable-version set beyond advisory metadata; merely finding a candidate does not prove the older version is affected.
Read the implicationsThe consequential AI-security question is how automated exploit migration changes the reach of a known bug across unlisted software history.
Limits and uncertaintyFull account
The benchmark sampled known Python web-application CVEs, confirmed original exploits and then tested older releases that vulnerability databases did not identify as affected. Agents were asked to migrate PoCs when the old source, API or dependency structure made a direct replay ineffective.
Researchers verified 149 older affected historical releases across 45 already-known CVEs; 22 reused the original PoC unchanged while 127 required migration. In selected agent evaluations, more than half of migration tasks succeeded when sufficient exploit and environment evidence was supplied.
These are missing vulnerable version ranges for previously known CVEs, not 149 previously unknown zero-days or 149 observed production compromises. Unresolved testing results must not be counted as unaffected versions.
Reported figures and denominator. 149 newly verified affected historical releases, 45 existing CVEs, 22 unchanged and 127 migrated PoCs; not zero-day or victim counts.
The consequential AI-security question is how automated exploit migration changes the reach of a known bug across unlisted software history.
Mechanism and trust boundary
- 01
Starting condition
An agent has access to a known exploit and older software versions whose API/implementation drift affects applicability.
Boundary: precondition and attacker capability
- 02
Trust boundary crossed or evaluated
Researchers verified 149 older affected historical releases across 45 already-known CVEs; 22 reused the original PoC unchanged while 127 required migration. In selected agent evaluations, more than half of migration tasks succeeded when sufficient exploit and environment evidence was supplied.
Boundary: mechanism
- 03
Observed consequence
An execution-checked migrated PoC extends the effective vulnerable-version set beyond advisory metadata; merely finding a candidate does not prove the older version is affected.
Boundary: observed effect and limitations
Claims & evidence
CLM-NOV15-CROSSPLOIT-MECHThe benchmark sampled known Python web-application CVEs, confirmed original exploits and then tested older releases that vulnerability databases did not identify as affected. Agents were asked to migrate PoCs when the old source, API or dependency structure made a direct replay ineffective. Researchers verified 149 older affected historical releases across 45 already-known CVEs; 22 reused the original PoC unchanged while 127 required migration. In selected agent evaluations, more than half of migration tasks succeeded when sufficient exploit and environment evidence was supplied.supported
Basis: reported finding
- supportsCrossPloit benchmark: agent PoC migration across old library versionsprimary disclosure
CrossPloit paper, benchmark construction, 149 historical releases, 127 migration tasks and experimental results
CLM-NOV15-CROSSPLOIT-SCOPEThese are missing vulnerable version ranges for previously known CVEs, not 149 previously unknown zero-days or 149 observed production compromises. Unresolved testing results must not be counted as unaffected versions. The consequential AI-security question is how automated exploit migration changes the reach of a known bug across unlisted software history.supported
Basis: reported finding
- supportsCrossPloit benchmark: agent PoC migration across old library versionsprimary disclosure
Primary disclosure: results, scope and limitations
Implications
An execution-checked migrated PoC extends the effective vulnerable-version set beyond advisory metadata; merely finding a candidate does not prove the older version is affected.
Controls and mitigations
- Verify advisory affected-version ranges with executable PoCs and keep inconclusive outcomes explicit.
- Separate proof migration under test authorization from exploitation of live systems; maintain historical library and deployment inventory.
Unknowns and contradictions
- The consequential AI-security question is how automated exploit migration changes the reach of a known bug across unlisted software history.
Sources and citation
Material revision history
- Nov 15, 2026 · Canonical change recorded · new in release · revision 250