When Vulnerability Discovery Shifts from Expert Hours to Model Inference: The Economic Balance of Security is Shifting

A recently public case warrants close attention in the security industry: a frontier AI model launched a specialized cybersecurity version, released only to strictly vetted customers. This version reproduced 1,724 historical vulnerabilities in mainstream open-source operating system kernels, produced 948 directly usable verification exploit snippets, and achieved over 5x the performance of previous models. It also discovered 81 unknown critical vulnerabilities in a major browser—promptly patched by the maintainers—and submitted 304 severe defects with remediation suggestions to open-source maintainers. The model’s internal cybersecurity capability level was escalated to critical.

What is more alarming than the volume of vulnerabilities is the shift in the capability chain. The model doesn’t just point out suspicious code; it generates the trigger conditions, verification paths, remediation suggestions, and verification exploit snippets together. Consequently, the marginal cost of vulnerability discovery has been drastically reduced. The public case notes a total compute cost of about $1.21 million, translating to just $1,279 per valid vulnerability. In contrast, traditional bug bounties for vulnerabilities of similar quality typically range from $20,000 to $70,000. A critical vulnerability used to represent weeks of mental labor by a security researcher; now, it increasingly resembles a scaled model inference expense.

This is good news for defenders, but it also signals a restructuring of offensive and defensive economics. The question is no longer “Can AI find vulnerabilities?” but “When vulnerability discovery and verification become low-cost, batch-processable, and assembly-line capabilities, how do enterprises rebuild their security operations?”

1. The Real Leap: From Vulnerability Discovery to Exploit Generation

Traditional vulnerability discovery roughly involves five stages: locating, triggering, reproducing, impact verification, and remediation. Manual research relies heavily on experience, while automated fuzz testing depends on seed quality and coverage. With the introduction of Large Language Models (LLMs), the most significant change isn’t replacing a single stage but simultaneously lowering the “cognitive cost” across multiple stages.

First, code comprehension costs decrease. Models can perform cross-file and cross-function data flow tracking across large codebases, linking “suspicious inputs—boundary conditions—dangerous operations” into candidate paths. Second, trigger condition generation costs decrease. Models can generate minimized test inputs based on candidate paths, reducing the time spent manually constructing samples. Third, verification and reporting costs decrease. Models can structure reproduction results into formal reports, providing patch drafts and regression test recommendations.

However, it must be emphasized: discovering a vulnerability and generating a runnable exploit snippet are two different risk levels. The former improves defenders’ triage efficiency; the latter means vulnerabilities are more easily confirmed and weaponized. The “948 directly usable verification exploit snippets” in the public case drew attention precisely because the capability loop advanced from “knowing there is a problem” to “proving it is exploitable.” This is why access controls, sandbox isolation, and output auditing must be upgraded simultaneously.

2. The Cost Breakdown Behind the $1,279 Figure: Not Simply Equivalent to Bounty Prices

The contrast between $1,279 per vulnerability and traditional bounties of $20,000 to $70,000 is striking. However, a rigorous analysis requires distinguishing the cost scopes.

Model inference cost is a marginal compute cost, typically covering computation, scheduling, storage, and partial verification overhead. Bug bounties, conversely, are market incentive prices that include scarcity premiums, research risks, quality guarantees, and long-term ecosystem incentives. The two are not strictly comparable. Still, the trend is clear: when candidate vulnerabilities can be batch-generated, automatically reproduced, and deduplicated, the discovery cost per valid vulnerability will drop significantly.

A simplified cost model can express this:

Cost per valid vulnerability = (Compute cost + Data & annotation cost + Isolated verification cost + Manual review cost + Remediation collaboration cost) ÷ Number of valid and reproducible vulnerabilities

Here, manual review costs cannot be omitted. Model outputs may contain false positives, duplicate reports, environmental dependency errors, and irreproducible conclusions. Without manual review, low costs translate into low-quality reports, increasing the burden on maintainers. The true cost advantage comes from the combination of “model batch screening + sandbox automated verification + manual review of high-value results only,” rather than treating model outputs as final vulnerability reports.

3. Defender Benefits: Scaling Detection and Remediation

For defenders, the most direct benefit is scaled detection. Enterprises can perform batch scans on code repositories, dependencies, configurations, and historical changes within authorized scopes, prioritizing high-risk paths. Models can also translate vulnerability descriptions into remediation suggestions, patch drafts, and regression test cases, shortening the pipeline from discovery to fix.

Yet, defenders cannot treat “the model says there’s an issue” as a final conclusion. A four-layer verification mechanism is recommended:

1. Authorization Boundary Verification: Confirm target assets, code scope, and testing windows are fully authorized.
2. Sandbox Reproduction Verification: Execute non-destructive verification snippets in an isolated environment and log environmental differences.
3. Deduplication and Prioritization: Sort by reachability, impact scope, exploitation conditions, and business exposure.
4. Remediation and Regression Loop: Submit patches to code review, cover trigger paths with regression tests, and feed fix results into metrics.

Special attention must be paid to supply chain remediation bottlenecks. Even if a model discovers numerous defects, maintainers’ capacity remains limited. If reports lack minimal reproduction steps, impact descriptions, and remediation suggestions, open-source maintainers will be overwhelmed by low-cost reports. Therefore, enterprises should adopt “reducing maintainer burden” as an output standard: reports are more valuable when they are highly structured, stably reproducible, and reviewable as patches.

4. Attacker Risks: Plummeting Costs if Access Controls are Breached

This capability is highly dual-use. While defenders gain a low-cost mining tool, attackers who bypass access controls can acquire vulnerability clues and verification paths at equally low costs. More dangerously, attackers gain not just a single vulnerability, but an entire pipeline from candidate locating to exploit generation.

Therefore, capability providers should establish at least the following guardrails: strict customer vetting and usage declarations; least privilege and tenant isolation; secondary approvals for sensitive tasks; output desensitization and export restrictions; comprehensive audit logging; anomalous behavior detection; offline isolated environments; rate limiting; and prohibiting the direct use of verification snippets on unauthorized targets. Model weights, evaluation samples, cached results, and logs are also highly sensitive assets; protecting just the inference interface is insufficient.

5. The Arms Race: Remediation Speed Over Model Speed

Another evaluation disclosed the same day showed that the model’s iteration speed increased to 1.9x its original rate, with 80% of the acceleration coming from AI self-assisted R&D rather than simply scaling compute. This means capability growth may outpace governance system development. The economic balance between offense and defense has been pried open: discovering vulnerabilities is cheaper, verifying them is cheaper, and generating remediation suggestions is cheaper. However, patch review, business regression, supply chain collaboration, and accountability will not automatically become cheaper.

For defenders to win, it’s not about who scans the most, but who fixes the fastest, verifies the most accurately, and has the fewest recurrences. We recommend focusing on metrics such as: valid vulnerability rate, false positive rate, verification pass rate, mean time to remediate (MTTR), recurrence rate, cost per valid vulnerability, critical asset coverage, and remediation suggestion adoption rate. Metrics should evaluate both “discovery efficiency” and “remediation quality,” otherwise organizations risk falling into an illusion of report volume prosperity.

6. A Reusable Pipeline: Enclosing Model Capabilities in a Controlled Loop

Below is an abstract pipeline pseudocode. The focus is on authorization, isolation, auditing, and manual review, rather than specific attack implementations:

Input: Authorized code snapshot, build configuration, asset scope, testing window
Candidate Set = Static Rule Scan(Code Snapshot) + Change Impact Analysis(Code Snapshot)
Candidate Set = Candidate Set ∪ Fuzzing Seed Scheduler(Build Configuration)
For each candidate:
    If NOT in authorization scope:
        Discard and log reason
    Trigger Condition = Model Generates Minimal Trigger Condition(Candidate)
    Reproduction Result = Isolated Sandbox Execution(Trigger Condition)
    If NOT reproduced:
        Mark as pending review false positive
    Verification Snippet = Model Generates Non-destructive Verification Snippet(Candidate, Reproduction Result)
    Patch Draft = Model Generates Patch Draft and Regression Test Suggestions(Candidate)
    Manual Review(Candidate, Reproduction Result, Verification Snippet, Patch Draft)
    Write to Audit Log(Candidate, Reproduction Result, Verification Snippet, Patch Draft)
Output: Deduplicated and sorted vulnerability reports, remediation suggestions, and metric data

The core of the pipeline is not “automating everything,” but placing the model in a controlled phase: candidate generation, sandbox verification, and report compilation can be automated. However, authorization judgments, risk grading, patch merging, and external disclosure must retain human accountability.

7. Conclusion: Security Victory Will Not Come Automatically

Low-cost vulnerability discovery presents a rare opportunity for capability equalization for defenders, but it also marks the beginning of risk equalization. The $1,279 per vulnerability cost indicates that vulnerability discovery is shifting from expert-hour intensive to model-inference intensive. Yet, vulnerability remediation, supply chain collaboration, and security governance still rely on policies, processes, and people.

The real watershed lies in whether enterprises can build a closed loop of “discovery—verification—remediation—metrics” during this window of cost reduction, implement strict access controls for highly sensitive capabilities, and transform model outputs into high-quality reports that maintainers are willing to accept. Achieving this turns low-cost discovery into a security victory. Failing to do so will only make it an accelerator for a new round of offense-defense arms races.