ARGUS · Cross-case synthesis · Preliminary

Public assurance stops one link short of the release decision

Across three frontier developers, we can inspect substantial published safety evidence, but we cannot fully reconstruct the final connection between that evidence and the decision to release.

3 DEVELOPERS · 3 CYBERSECURITY RELEASES · SEPT–OCT 2026 · NOT EXTERNALLY PEER REVIEWED

The assurance chain, traced through each case

Five stages: capability assessments, safeguards and testing, adverse findings and remaining risk, risk-acceptance reasoning, release decision. For GPT-6 Astra, Claude Opus 5.5 and Gemini 4 Argon the published evidence carries through the first three stages; at risk-acceptance reasoning each case records an open gap (Proof-1 G2 and G3, Proof-2 G-007, Proof-3 G-002), so the connection to the release decision is partly or not independently inspectable.

PROOF-1AstraPROOF-2Opus 5.5PROOF-3Argon01CapabilityassessmentsPUBLISHED EVIDENCE02Safeguardsand testingPUBLISHED EVIDENCE03Adverse findingsand remaining riskPUBLISHED EVIDENCE04Risk-acceptancereasoningNOT INSPECTABLE05ReleasedecisionPUBLICG2 · G3G-007G-002
Conceptual synthesis of the three assurance cases, not a literal shared decision procedure. Each case keeps its own boundary, vocabulary and date; the stage placement of its gaps is ARGUS’s synthesis.Stages 1 to 3 are not gap-free: every case records other open links there. The chart isolates the one link all three share.

The break, case by case

Where each reconstruction loses the thread

Gap statements are quoted from each case’s filed graph. Select a trace in the chart to jump to its case.

PROOF-1 · OPENAI

GPT-6 Astra

Cybersecurity · public release · evidence cutoff 3 Sep 2026

G2

Public measurements are not connected to an inspectable deployment-level residual severe-harm assessment.

G3

Critical-cyber safeguard/security acceptance standards and the launch-specific efficacy thresholds are not sufficiently public to replay the decision.

Both unchanged at the 9 Oct 2026 public re-check (Proof-1.1). 9 open links in the case.

Open the Proof-1 graph

PROOF-2 · ANTHROPIC

Claude Opus 5.5

Cybersecurity · release baseline 22 Sep 2026

G-007

General FCF process to model-specific decision evidence

The inspected corpus does not disclose an Opus 5.5 cyber acceptance rationale connecting capability, adverse results, fallback and the temporary margin to that condition, or model-specific monitoring effectiveness.

Open in the public record examined. 8 open links in the case.

Read the Proof-2 gaps

PROOF-3 · GOOGLE DEEPMIND

Gemini 4 Argon

Cyberoffensive misuse · broad rollout 13 Oct 2026

G-002

Safety-case and acceptance bridge

The complete model-specific cyber safety case, residual-risk calculation and governance acceptance decision linking unmitigated 16.7x uplift to below-CCL deployment risk are not publicly inspectable.

Open in the public record examined. 7 open links in the case.

Read the Proof-3 gaps

That is the recurring last-mile problem in public frontier AI assurance.

What this does not show

It does not establish that these companies lacked internal evidence, concealed failures or made incorrect decisions. Nor have three cases established an industry-wide prevalence rate.

  • Gap counts are link judgements within each case, not safety scores, and are not comparable across cases.
  • All three are documentary reconstructions of public sources in one risk domain, cybersecurity.
  • None has been externally peer reviewed.

Preliminary synthesis, 14 October 2026. Independent research by Automa Dynamics; not affiliated with or endorsed by OpenAI, Anthropic or Google.

About ARGUS