{"snapshot":{"id":"civ-20261009-ai-verification-bottleneck","as_of":"2026-10-09 04:25:30.000","summary":"AI-assisted vulnerability discovery has reached a scale where human verification and remediation are visibly lagging candidate generation in a large open-source security program. Anthropic reports 29,439 candidate findings and 6,123 externally triaged findings as of October 2, with 5,674 confirmed valid; maintainers have already patched a subset and independent project records confirm fixes originating from the program. The shift is real but bounded: it does not show that all AI-generated findings are correct, that remediation is automated, or that human verification has become the dominant bottleneck across every domain.","cross_layer_changes":["Model-driven vulnerability discovery is producing security candidates at higher throughput than expert review in a large production-adjacent pipeline.","Open-source maintainers can opt to receive unreviewed machine findings faster, explicitly trading review latency against 0-positive and prioritization risk.","Human security expertise moves downstream toward verification, prioritization, disclosure and safe repair rather than disappearing from the loop."],"dependencies":["models + software analysis -> vulnerability candidate generation","human security review + maintainers -> validation, prioritization and remediation"],"constraints":["human triage capacity","maintainer bandwidth","0 positives and severity errors","safe patch verification","disclosure latency"],"emerging_capabilities":["recurring model-generated open-source vulnerability scanning","machine-produced reproducers and candidate patches"],"possible_bottlenecks":["expert verification throughput","maintainer remediation capacity","secure deployment of fixes"],"uncertainties":["generality beyond Anthropic's program","long-run precision of unreviewed scans","fraction of candidates that become exploitable, accepted and patched vulnerabilities"],"mapping_mode":"contemporaneous","reconstructed_at":null,"created_at":"2026-10-09T04:28:55.056Z","updated_at":"2026-10-09T04:28:55.056Z"},"claims":[{"id":"claim-ai-vuln-verification-bottleneck-20261009","statement":"Anthropic's open-source vulnerability program provides concrete evidence that model-generated vulnerability discovery can outpace available human triage capacity, with external review confirming a large share of reviewed findings while remediation remains substantially downstream.","epistemic_status":"supported_inference","confidence":0.96,"civilization_snapshot_id":"civ-20261009-ai-verification-bottleneck","layer_state_snapshot_id":"state-human-20261009-verification-throughput","competing_interpretation":"The throughput gap may be specific to Anthropic's program design and disclosure process rather than a general human-AI bottleneck.","uncertainty":"Cross-provider replication, representative unreviewed precision and end-to-end remediation latency remain incomplete.","mapping_mode":"contemporaneous","reconstructed_at":null,"created_at":"2026-10-09T04:28:55.056Z","updated_at":"2026-10-09T04:28:55.056Z"}],"layer_states":[{"id":"state-human-20261009-verification-throughput","layer_id":"layer-human","as_of":"2026-10-09 04:25:30.000","summary":"Human verification capacity is becoming a measurable bottleneck in AI-accelerated software security. Anthropic's open-source pipeline generated 29,439 candidate findings while only 6,123 had reached external human triage by October 2; 5,674 of those were confirmed valid. Independent maintainer release records also show real fixes from the program. The evidence is strong for a verification-throughput constraint, but it is concentrated in one vendor-led cybersecurity program and does not establish the same bottleneck across all AI-assisted knowledge work.","dependencies":["layer-models","layer-software","layer-open-source","layer-research"],"constraints":["expert triage capacity","maintainer attention","safe patch validation","disclosure coordination","severity calibration"],"emerging_capabilities":["AI-generated vulnerability reports with reproducers and candidate patches","high-throughput machine-assisted security discovery"],"underdeveloped_components":["automated trustworthy triage","safe automated remediation","maintainer throughput at discovery scale","independent cross-provider replication"],"uncertainties":["how representative Anthropic's pipeline is of other domains and providers","how raw unreviewed scanner precision changes outside selected high-severity samples","how quickly disclosed findings convert into deployed fixes"],"confidence":0.95,"previous_snapshot_id":"state-human-20261004-deployment-capacity","mapping_mode":"contemporaneous","reconstructed_at":null,"created_at":"2026-10-09T04:28:55.055Z","updated_at":"2026-10-09T04:28:55.055Z","layer_slug":"human","layer_name":"Human"}],"machine_reading":{"recommended_start":"/machine/context.json","recommended_reading_order":["/machine/context.json","/machine/current-world.json"],"complete_current_world":"/machine/current-world.json","public_mcp":"/mcp-public","scope":"civilization_head","note":"This endpoint is not the complete current-world projection. Begin with the machine-reading contract, then read the complete current world."}}