OpenAI Enlists Cathedral Systems to Probe Its Own Defenses With AI

On November 2, 2026, OpenAI disclosed an internal defensive initiative enlisting advanced artificial intelligence models to identify software vulnerabilities across company systems before external actors exploit them. The program pairs internal security engineering with external researchers to challenge defenses and test remediations.

Under the revised operating model, technical teams incorporate frontier language systems directly into exploratory penetration testing rather than relying solely on human code auditing. OpenAI partnered with cybersecurity startup Cathedral Systems, deploying external research personnel alongside internal staff to stress-test core application environments. The company previously allocated approximately 25% of its product engineering staff to an internal security sprint powered by Codex, applying model capabilities to inspect codebases and stage reproduction sandboxes. The current expansion broadens those testing efforts, although OpenAI has not disclosed the specific models, timelines, or systems included in the collaboration.

External security researchers will operate alongside OpenAI teams using the company's frontier models to locate and stress-test vulnerabilities.

Concurrent with the technical sprint, OpenAI restructured its financial reward schedules for independent security contributors, as documented by Help Net Security. Baseline compensation for critical flaws in primary consumer applications rose to $30,000, while exceptional structural discoveries now command base allocations of up to $100,000. Through newly integrated performance multipliers, exceptional vulnerability submissions can yield payouts reaching $171,925, with an announced maximum ceiling of $172,000 when combined with bonus pools. The dual-track strategy reflects a broader industrial recognition that securing complex model ecosystems requires both automated machine evaluation and disciplined human testing.

Autonomous Vulnerability Hunting Across Production Systems

The partnership with Cathedral Systems marks an operational effort to apply model capabilities directly to defensive vulnerability research.

In previous defensive cycles, OpenAI utilized specialized code models to examine repository commits, isolate anomalous logic paths, and synthesize candidate security patches for human review. According to documentation published by OpenAI, automated agents functioned within isolated execution environments to confirm whether hypothesized exploits could execute arbitrary instructions. Engineers validated reproducible test cases before committing structural remediations into production branches, accelerating remediation velocity across distributed web services. The integration of Cathedral Systems introduces external operator viewpoints, verifying whether autonomous inspection tools identify logic oversights missed during standard code evaluations.

OpenAI operates a separate Safety Bug Bounty program established in March 2026 to track model behavior and jailbreaks rather than infrastructure software flaws.

Technical personnel configure inspection workflows to evaluate system resilience and discover potential regression flaws across development branches. Evaluating software resilience through automated verification loops allows engineering teams to remediate structural weaknesses before hostile scanning operations discover similar execution vectors, establishing rigorous red-teaming practices across the organization. The approach contrasts with traditional manual penetration tests, which often conclude after episodic assessments of fixed software baselines. However, OpenAI has not published performance metrics or operational timelines demonstrating how quickly vulnerabilities are resolved under the new arrangement.

The Anthropic Precedent and Enterprise Software Exposure

Cathedral Systems gained industry prominence after conducting a security research exercise that accessed internal environments at Anthropic.

On October 26, 2026, an investigative report published by The Washington Post revealed that Cathedral Systems researchers accessed internal systems at Anthropic in under 12 hours during a security assessment. Rather than breaching frontier model weights or proprietary algorithmic architectures, the research team exploited vulnerabilities embedded within standard third-party enterprise tools. The compromised surfaces granted access to internal meeting transcripts, employee salary records, and messaging channels discussing advanced model capabilities. Anthropic confirmed the intrusion, decommissioned the vulnerable business applications, and notified federal authorities regarding the unauthorized access.

Anthropic stated that researchers never accessed customer conversations with Claude, model weights, or core algorithmic training data.

Anthropic criticized Cathedral for overstating the national security implications of the exercise and asserted that researchers failed to strictly follow testing and disclosure rules. The startup, valued at $1.4 billion after raising $160 million, was co-founded by former DOGE staffers, including Gavin Kliger, who previously served as the Pentagon's chief data officer. The episode illustrated how supply-chain dependencies in commercial software create entry corridors into frontier technology laboratories. Contracting the firm provides OpenAI with specialized expertise in testing non-core software perimeters against automated intrusion methodologies.

Economic Multipliers and Bug Bounty Restructuring

Reward midpoints across medium, high, and critical vulnerability tiers increased more than fourfold under OpenAI's revised compensation policy.

Under policy revisions detailed by OpenAI, reward midpoints across medium, high, and critical vulnerability tiers increased more than fourfold on average. The organization implemented three compounding bonus mechanisms: report quality multipliers, a five-tier researcher loyalty system, and targeted priority campaign allocations. The published payment schedule calculates that a $100,000 base bounty expands to $171,925 when high-fidelity documentation, exploit demonstrations, and loyalty multipliers align, under a maximum ceiling of $172,000. The company also announced retroactive adjustments for qualifying submissions processed over the preceding 30 days.

The retroactive policy applies new reward scales to pending reports and provides supplemental compensation for qualifying bounties awarded within the prior 30 days.

Independent bug bounty programs provide essential verification that internal testing regimes cannot replicate in isolation. By increasing rewards for high-severity findings, the updated compensation table routes elite security researchers toward responsible disclosure channels rather than gray-market brokers. The compensation schedule distinguishes technical vulnerabilities affecting platform infrastructure from behavioral model jailbreaks, which remain governed under a separate program established in March 2026. Systematic verification of credential isolation, API gateways, and authorization barriers provides empirical protection across the broader commercial ecosystem.

Defensive Acceleration and the Shift Toward Machine-Speed Patching

The integration of AI models into vulnerability research reflects an industry effort to accelerate defensive response times.

Traditional cybersecurity paradigms rely on periodic human reviews that struggle to match the velocity of automated scanning tools operated by hostile groups. Integrating models into defensive code loops allows software developers to simulate adversarial behaviors before external intrusions occur. When testing systems generate reliable reproduction scripts in sandboxed environments, engineering teams can validate security patches efficiently. This continuous verification model reduces the operational window during which unpatched flaws remain exposed across public cloud clusters and developer endpoints.

OpenAI has not disclosed performance metrics or verification baselines evaluating the effectiveness of model-assisted vulnerability discovery.

OpenAI has yet to publish empirical telemetry, scope boundaries, or audit timelines detailing the specific model architectures slated for deployment alongside Cathedral Systems. The defensive sprint establishes a baseline test for whether artificial intelligence models can secure commercial technology clusters as effectively as external research teams probe enterprise networks. Evaluating both third-party SaaS applications and internal platform services remains necessary to prevent unauthorized access across expanding corporate footprints. Machine learning security will increasingly depend on disciplined algorithmic auditing alongside continuous human oversight.