Lifting Filters Reveals the Offensive Capabilities of AI Models
The American artificial intelligence lab Anthropic has restructured its access framework for cybersecurity specialists through its Cyber Verification Program. By unifying this program with experiments conducted under Project Glasswing, the company now offers three distinct access tiers: operational defence, red-team offensive simulation, and restricted critical infrastructure access. This segmentation applies to its most advanced models, including Claude Opus 5.5, Claude Sonnet 5.5, and the experimental Mythos family.
Alongside this announcement, Anthropic published findings from CyScenarioBench, a standardized benchmark designed to assess a model's ability to plan and execute multi-step intrusion operations under realistic constraints. The results reveal a sharp contrast: across 50 interactive offensive scenarios, the standard public model was blocked on the very first prompt in every trial. In the intermediate defensive tier, 46 of the 50 queries encountered interruption mechanisms. Under Red Team Access, however, no blocks occurred, and Claude Opus 5.5 completed 34 of the 50 missions, matching the success rate of a test run without any safety guardrails.
This algorithmic transparency comes alongside significant evidence regarding software ecosystem vulnerabilities. Between April and July 2026, Project Glasswing industry partners identified over 129,000 confirmed software flaws, with more than 33,000 rated critical or high severity. As Bloomberg Television reported, JPMorgan Chase chief executive officer Jamie Dimon noted that cybersecurity risks have expanded dramatically with frontier models that can probe source code at unprecedented speed.
The Dual-Use Dilemma and the Opacity of Commercial Guardrails
These findings clearly demonstrate the dual-use nature of frontier models. The same neural networks used to audit code, reverse-engineer malware, and generate patches can also identify zero-day vulnerabilities and orchestrate attack vectors. To curb abuse without paralyzing defensive teams, providers rely on probabilistic classifiers and moderation filters upstream of user prompts.
However, this approach introduces substantial operational hurdles, as documented by specialized outlets like The Hacker News and Help Net Security. Automated classifiers frequently struggle to separate legitimate vulnerability research conducted by an internal team from malicious reconnaissance, causing recurring false positives that disrupt standard development workflows. To manage these rigid controls, vendors impose burdensome corporate identity verification processes and mandate ongoing logging of prompt data to monitor for misuse, raising immediate business privacy concerns.
Regulatory agencies have expressed similar concerns. In recent statements regarding advanced artificial intelligence models, the Canadian Centre for Cyber Security emphasized that automated vulnerability discovery drastically compresses the time defenders have to remediate flaws. Meanwhile, the OWASP project lists excessive agency among the primary security risks for large language model applications: assigning diagnostic tools or system access to autonomous agents without strict sandboxing exposes organizations to unintended drift or guardrail bypasses.
The ProductivIA Perspective: Observing and Comparing with GoIA
For corporate organizations, these technical developments highlight a strategic risk: becoming locked into a closed ecosystem where moderation rules, filtering criteria, and actual model capabilities are determined unilaterally by a single foreign provider. Lacking the ability to test alternative computational architectures, organizations cannot confirm whether an unanswered prompt reflects an overly aggressive filter, a reasoning limitation, or an architectural constraint.
To provide this needed clarity, ProductivIA incorporates the GoIA application. Designed as a multi-model workspace within a unified browser interface, GoIA allows enterprise teams to query multiple distinct models simultaneously, such as Anthropic, OpenAI, Mistral, and Quebec-hosted sovereign model Matania, from a single prompt.
Rather than accepting answers from a single ecosystem without verification, managers and engineers can compare model outputs side by side during complex queries, process audits, or document reviews. This empirical benchmarking clarifies provider-specific biases, rejection thresholds, and reasoning depths. Within an enterprise environment subject to Quebec Law 25 compliance obligations, this diversified orchestration prevents vendor lock-in: organizations keep control over their technology stack and can redirect sensitive queries to sovereign infrastructure without altering workflows or writing custom integration code.
Looking Ahead
As artificial intelligence laboratories build tiered, closely monitored access models, how will public and private organizations protect their independent audit capabilities when using closed systems? Adopting neutral evaluation platforms and diversifying model providers represent the first line of defence against technological opacity.