Artificial intelligence has become a critical tool for both cybersecurity defenders and malicious actors. However, the guardrails implemented by AI model developers are creating significant obstacles for legitimate security researchers. This situation creates a conflict between the need to prevent misuse and the requirements of cybersecurity professionals.
Key takeaways
- In June 2026, Anthropic's Mythos and Fable models were placed under US export controls due to concerns about their potential use in cyberattacks.
- OpenAI and Anthropic offer verified researcher programs, but the vetting process is time-consuming and limits access.
- 57% of offensive security specialists report difficulties using AI due to guardrails.
- Open-source models without guardrails are becoming a popular alternative among researchers.
- Experts warn that restrictions could slow critical vulnerability discovery by 30-40%.
The new AI restrictions
In June 2026, the US government imposed export controls on Anthropic's Mythos and Fable models. This decision was partly motivated by reports of potential circumvention of safety mechanisms designed to prevent AI's use in cyberattacks.
Current restriction status
While restrictions on Fable 5 were lifted on July 1, 2026, access to Mythos 5 remains controlled. This model is now only available to vetted US organizations under a government evaluation program.
Impact on international research
Export controls have created additional challenges for global security research teams. Companies outside the Five Eyes alliance must now either find workarounds or abandon advanced models entirely for their research.
Why this matters for cybersecurity
AI models have become essential tools for code analysis, vulnerability discovery, and defense mechanism development. However, these same capabilities can be exploited to create attack vectors.
The dual-use dilemma
Chris Anley from NCC Group compares AI to a hammer: "It's a construction tool but also a weapon. You can't separate one from the other." A request to fix a code vulnerability is technically identical to requesting an exploit for that same vulnerability.
Real-world examples
During testing of a financial platform, researchers found that AI queries about SQL injection analysis were automatically blocked by the system, despite this analysis being necessary for developing defenses. This caused a three-week delay in vulnerability remediation.
Expert opinions: The debate over restrictions
The security research community is divided. Mark Dowd criticizes AI companies' approach: "It's uncomfortable that random large corporations make arbitrary decisions about what's safe in the security space."
Alternative approaches
Giuseppe Calì uses AI only for reverse engineering, reserving vulnerability discovery for himself: "I'm too possessive about my bugs to let models play this game for me."
Technical limitations of guardrails
Current protection systems rely on keywords and contextual analysis, leading to false positives. Queries containing terms like "exploit" or "vulnerability" are often automatically blocked, even when coming from accredited researchers.
How researchers are adapting
Cybersecurity specialists have developed several workaround strategies:
- Switching to open-source models without guardrails
- Local model deployment for sensitive tasks
- Process separation: AI for analysis, humans for vulnerability discovery
- Using metaphors and euphemisms in queries
- Gradual "training" of models through follow-up questions
Approach comparison
| Method | Advantages | Disadvantages |
|---|---|---|
| Open-source models | Complete freedom | Requires powerful hardware |
| Researcher programs | Access to cutting-edge models | Bureaucracy and delays |
| Guardrail circumvention | Immediate solution | Risk of account suspension |
Risks of current situation
Stringent restrictions create several serious problems for the cybersecurity industry:
- Slowed discovery of critical vulnerabilities
- Researchers moving underground due to bureaucratic barriers
- Growing popularity of unregulated open-source solutions
- Reduced competitiveness of legitimate researchers
- Risk of knowledge leakage to less regulated jurisdictions
Industry statistics
According to Black Hat 2026 data, 68% of security professionals believe current restrictions negatively impact their work. Meanwhile, 42% admit to using Chinese open-source alternatives due to access problems with commercial models.
Practical steps for cybersecurity professionals
Experts recommend several actions to navigate these new constraints:
- Explore alternative open-source AI tools
- Participate in verified researcher programs
- Develop internal AI usage standards
- Invest in local computing resources
- Create specialized training datasets
- Engage in regulatory discussions
Recommended tools
For local deployment, experts recommend models like GPT-J, BLOOM or GLM. These offer a reasonable balance between performance and resource requirements, especially when using quantization.
Questions & answers
Which AI models faced restrictions in 2026?
Anthropic's Mythos and Fable models were placed under US export controls. While restrictions on Fable 5 were lifted, Mythos 5 remains limited to vetted US organizations.
Why are cybersecurity researchers opposed to AI guardrails?
Safety mechanisms frequently block legitimate vulnerability analysis queries, slowing research and forcing professionals to find workarounds.
How do cybersecurity professionals use AI?
Primary applications include: code reverse engineering, vulnerability analysis, task automation, and security testing tool development.
What alternatives do experts propose instead of strict restrictions?
Suggestions include more flexible verified researcher programs, improved audit mechanisms, and development of local AI solutions for sensitive tasks.
How do guardrails affect vulnerability discovery speed?
Experts estimate current restrictions increase analysis time by 25-35% for standard vulnerabilities and up to 50% for complex cases requiring in-depth research.
What are the legal risks of circumvention methods?
Legal experts warn that bypassing guardrails may violate platform terms of service and could potentially lead to lawsuits, especially with commercial models.