When Bharti Airtel's CTO Randeep Sekhon stood before reporters at a COAI industry event and admitted his team is now in active talks with vendors like Ericsson, Nokia, and Cisco to assess risks flagged by an AI model, it signaled something that security practitioners have been anticipating — and quietly dreading. Airtel is working with its global suppliers and technology partners to identify and fix vulnerabilities that advanced AI models can detect more effectively than traditional systems. The catalyst? Anthropic's Claude Mythos, an AI model currently available to a limited set of companies, can identify cybersecurity vulnerabilities in digital infrastructure — raising concerns that, if accessed or misused by malicious actors, it could increase cyberattack risk across critical infrastructure sectors such as telecom and banking.
This is not a theoretical threat. This is what happens when AI capability leapfrogs the security architecture built over the past two decades. If you run a SOC, manage a telecom network, or sit in a CISO chair anywhere in the world, what Mythos represents demands your immediate attention.
What Claude Mythos Actually Does — and Why Telcos Are Nervous
Most public coverage of Mythos has focused on marketing superlatives. The technical reality is more sobering. During testing, Mythos Preview demonstrated the ability to identify and exploit zero-day vulnerabilities in every major operating system and every major web browser when directed by a user to do so. The vulnerabilities it finds are often subtle or difficult to detect — many of them ten or twenty years old, with the oldest being a now-patched 27-year-old bug in OpenBSD, an operating system known primarily for its security.
To understand why this matters for telecoms specifically, consider the attack surface. A tier-1 operator like Airtel runs proprietary vendor software embedded in BSS/OSS stacks, RAN infrastructure, and core network elements — none of which its own security team directly controls or can patch unilaterally. When Sekhon says "we don't do this, the software is owned by them," he is acknowledging a fundamental reality of telecom security: the attack surface belongs to third parties.
The Vendor Software Problem in Critical Infrastructure
Telecom operators are increasing coordination with vendors like Ericsson, Nokia, and Cisco given their core software and infrastructure reliance — vendors play a critical role in security, as telcos do not control the proprietary software embedded within these systems. This creates a classic supply chain security problem mapped directly to MITRE ATT&CK T1195 (Supply Chain Compromise). A vulnerability in a vendor-owned component can persist for years because no internal red team ever has access to the source code.
The Exploit Window Has Already Collapsed
The exploit window — the time between a vulnerability being known and being weaponized — has already collapsed from months to hours. Mythos doesn't just accelerate that. It eliminates it entirely for whoever has access. For telecom operators running 24/7 infrastructure with zero acceptable downtime, this changes every calculation in the patch management playbook.
Important: The assumption that "no CVE means no risk" is now definitively broken. Over 99% of the vulnerabilities Mythos has found remain unpatched. If a hostile actor develops equivalent capability, the entire "find, disclose, patch" model the security industry depends on becomes obsolete — no patch management cycle can close the gap in time.
The Dual-Use Dilemma: When the Defender's Tool Becomes the Attacker's Weapon
Mythos was restricted precisely because Anthropic recognized its offensive potential before its defensive promise. Mythos was able to weaponize a set of vulnerabilities it found in Firefox into 181 usable attacks; Anthropic's previous flagship model could only achieve two. That is not a linear improvement — it is a regime change.
Anthropic launched Project Glasswing as an urgent attempt to put these capabilities to work for defensive purposes, giving major technology, cybersecurity, and financial organizations early access to the model — committing up to $100M in usage credits for Mythos Preview and $4M in direct donations to open-source security organizations.
The problem, as any experienced threat intelligence analyst will tell you, is that controlled access programs are not containment. The Mythos model was reportedly accessed by a handful of users in a private online forum on the same day that Anthropic announced plans to release the model to a limited number of companies for testing — one of the members of the group was a third-party contractor for Anthropic. Insider access, not zero-day exploitation, was the vector. That is MITRE ATT&CK T1078 (Valid Accounts) applied at the AI supply chain level.
What Autonomous Exploit Generation Looks Like in Practice
Mythos Preview autonomously identified and exploited a 17-year-old remote code execution vulnerability in FreeBSD (CVE-2026-4747) that allows anyone to gain root on a machine running NFS — fully autonomously, with no human involved in either the discovery or exploitation after the initial request.
Engineers at Anthropic with no formal security training asked the model to find remote code execution vulnerabilities overnight, and woke up to complete, working exploits. The skills barrier — which previously kept sophisticated vulnerability research confined to elite red teams — has effectively dropped to near zero.
Airtel's Vendor Engagement Strategy: Right Response, But Is It Fast Enough?
Sekhon's response — talking to suppliers — is the correct first step. It maps to NIST SP 800-161 (Supply Chain Risk Management) and aligns with CIS Control 15 (Service Provider Management). But the pace of vendor engagement in enterprise telecom has historically been measured in quarters, not days. That cadence is now dangerously misaligned with the threat.
As AI systems continue to evolve, the balance between defensive and offensive cyber capabilities could shift — necessitating faster patch cycles, tighter vendor collaboration, and more adaptive security architectures.
What a Mature Response Looks Like
| Response Layer | Action Required | Framework Reference |
|---|---|---|
| Vendor Management | Contractual SLAs for AI-flagged CVE remediation | NIST SP 800-161, ISO 27036 |
| Threat Intelligence | Subscribe to Project Glasswing disclosures | CIS Control 17 |
| Patch Management | Compress patch cycles from 30-day to 72-hour for critical flaws | CIS Control 7, PCI DSS Req. 6 |
| Detection Engineering | Update SIEM rules for exploit-chaining behaviors (T1210, T1068) | MITRE ATT&CK |
| Incident Response | Tabletop exercises simulating AI-generated zero-day exploitation | NIST SP 800-61 |
Pro Tip: Don't wait for a CVE to be published before engaging your vendor. Mythos-class tools will identify vulnerabilities that never enter the CVE queue — especially in proprietary, closed-source telecom software. Establish a direct vulnerability disclosure channel with every tier-1 network vendor now, before you need it.
The Broader Industry Picture: From Firefox to 5G Core
Mozilla's experience provides the clearest data point available. Since February 2026, Mozilla's Firefox security team had been collaborating with Anthropic to scan the browser's codebase. An earlier phase leveraged Claude Opus 4.6, which identified 22 vulnerabilities — 14 of them high-severity — during a two-week engagement. Building on that, Mozilla applied Claude Mythos Preview to Firefox's codebase, resulting in 271 vulnerabilities patched in Firefox 150.
Palo Alto Networks also shared preliminary data from testing Mythos, noting that in terms of vulnerability discovery it accomplished the equivalent of a year's worth of pentesting in less than three weeks — with impressive vulnerability-chaining capabilities, combining medium- and low-severity issues into a critical exploit.
Now apply that calculus to a 5G core network. The attack surface is orders of magnitude larger, the proprietary code is not open for external review, and the regulatory consequences of a breach extend across GDPR, TRAI mandates, and critical infrastructure protection frameworks. A single chained exploit targeting a telecom OSS platform could expose subscriber data for hundreds of millions of users — a GDPR enforcement scenario that makes any Firefox patch cycle look trivial.
The Attack Surface Comparison
| Target Environment | Known Vulnerability Density | AI Discovery Advantage | Compliance Exposure |
|---|---|---|---|
| Web Browsers (Firefox) | High, well-audited | 271 new findings in one pass | Low-medium |
| Open-Source OS (OpenBSD, FreeBSD) | Medium, decades-old flaws hidden | 27-year-old bugs surfaced | Medium |
| Telecom Vendor Software (RAN, BSS/OSS) | Unknown — closed source | Largely unassessed | Critical — GDPR, TRAI |
| Enterprise Network Devices (Cisco, Nokia) | Medium, patched via advisories | Vendor-dependent disclosure | High — PCI DSS, HIPAA |
What Security Teams Must Do Right Now
The posture shift required here is architectural, not operational. Tweaking your existing vulnerability management program is insufficient when the threat model has fundamentally changed. Here is the strategic response framing:
- Accelerate third-party risk assessments — demand AI-augmented penetration testing from network vendors under existing contracts or renegotiated SLAs aligned with ISO 27001 Annex A.15
- Instrument for exploit chaining — update SIEM correlation rules to detect multi-stage exploitation patterns (T1068 privilege escalation chained with T1210 remote services exploitation) rather than single-indicator alerts
- Compress your patch SLA now — 30-day patch cycles were adequate when exploit windows measured in months; they are indefensible when Mythos-class tools eliminate that window entirely
- Engage Project Glasswing-equivalent programs — lobby your vendor ecosystem to adopt similar AI-driven defensive scanning under coordinated disclosure frameworks
- Conduct a supply chain tabletop — specifically simulate a scenario where a vendor-owned component contains an AI-discovered zero-day weaponized within 24 hours of discovery
Conclusion
What Airtel's CTO acknowledged publicly is what every telecom security architect already knows privately: the threat surface does not belong to them. The software that runs their network belongs to vendors operating on patch timelines designed for a slower threat era. Within six months, advanced AI models with deep cybersecurity capabilities will become commonplace — organizations that have not put appropriate safeguards in place will face an entirely new class of risk across their enterprise and critical infrastructure.
The Mythos disclosure is not a reason for panic. It is a forcing function. The same capability that autonomously found a 27-year-old OpenBSD vulnerability can be turned toward securing your infrastructure — but only if you move faster than the threat actors waiting for equivalent access. Start with your vendor contracts. Demand AI-augmented security assessments. Compress your patch cycles. The organizations that treat this moment as a compliance checkbox will regret it in 12 months.
Key Takeaways
- Claude Mythos can autonomously discover and weaponize zero-day vulnerabilities in major OS platforms and browsers — eliminating the traditional exploit window entirely
- Telecom operators face unique exposure because proprietary vendor software sits outside their direct security control, requiring urgent supply chain risk management under NIST SP 800-161
- Patch management SLAs must be renegotiated immediately — 30-day cycles are now a liability, not a standard
- MITRE ATT&CK T1195 (Supply Chain Compromise) and T1210 (Exploitation of Remote Services) should be primary threat model inputs for any telecom or critical infrastructure SOC
- Project Glasswing's controlled access model is a temporary containment measure, not a long-term solution — equivalent capability will proliferate
- Demand contractual AI-augmented penetration testing from every tier-1 network vendor before year-end
FAQ
Q: What exactly is Claude Mythos and why is it different from previous AI security tools?
Claude Mythos Preview is Anthropic's most capable AI model, purpose-evaluated for cybersecurity. Unlike earlier tools that flagged known vulnerability patterns, Mythos autonomously discovers previously unknown flaws — including a 27-year-old OpenBSD bug and a 17-year-old FreeBSD remote code execution vulnerability — and writes functional exploits without human guidance. The performance gap versus its predecessor is not incremental: where Claude Opus 4.6 produced two working Firefox exploits from hundreds of attempts, Mythos produced 181.
Q: Why is Airtel engaging vendors rather than fixing vulnerabilities directly?
Because telecom operators do not own or control the source code running their networks. The BSS/OSS platforms, RAN software, and core network elements are proprietary vendor products. Airtel cannot patch what it did not write — it can only pressure vendors to accelerate their own remediation cycles and share vulnerability intelligence upstream. This is standard supply chain risk management, but the urgency is now dramatically higher.
Q: Can Mythos-class AI be used by attackers, not just defenders?
Yes, and this is the central concern driving Anthropic's restricted access model. The same capability that finds vulnerabilities defensively can identify and weaponize them offensively. The model was already accessed by unauthorized users on the day of its announcement, which demonstrates that access control alone is not a containment strategy. Any organization running critical infrastructure should assume adversarial actors are pursuing equivalent capability.
Q: What compliance frameworks apply to this kind of AI-driven vulnerability risk?
Multiple frameworks are implicated. NIST SP 800-161 governs supply chain risk management. ISO 27001 Annex A.15 addresses supplier relationships. PCI DSS Requirement 6 mandates vulnerability management for cardholder data environments. GDPR Article 32 requires appropriate technical measures — and an AI-discovered zero-day in a telecom platform serving hundreds of millions of users would trigger breach notification obligations across every jurisdiction where subscribers reside.
Q: How should a SOC team update its detection posture in response to AI-generated exploits?
The primary shift is from single-indicator detection to behavior chain analysis. AI-generated exploits are more likely to chain low-severity vulnerabilities into critical exploits (a capability Mythos demonstrated explicitly). Update SIEM correlation rules to detect sequences involving T1068 (Exploitation for Privilege Escalation), T1210 (Exploitation of Remote Services), and T1059 (Command and Scripting Interpreter) within compressed timeframes. Also prioritize anomaly detection on vendor-managed network segments that your team cannot directly instrument
Enjoyed this article?
Subscribe for more cybersecurity insights.