Mythos, Glasswing 도대체 뭐길래

Overview of the discussion (Mythos vs. Glasswing / Glasswing meaning)

• The video explains “Mythos” as an AI foundation model by Anthropic, described as a more capable “next version” compared with prior Anthropic models such as Haiku / Sonnet / Opus.
• Mythos is not marketed as a cybersecurity-only model. It is positioned as general-purpose intelligence that becomes highly capable in cybersecurity when paired with the right pipeline.
• Because of potential risk and impact, Anthropic’s Mythos Preview is described as restricted-access, provided first to selected companies/institutions rather than the public.
• “Glasswing” is described as the name of an initiative/project focused on defensive vulnerability discovery—transparently identifying vulnerabilities and coordinating disclosure and fixes. The name is compared to a butterfly with transparent wings and few natural predators.

---

What Glasswing is intended to do (defensive pipeline + alliance)

Glasswing’s stated purpose is to use Mythos to:

• Pre-screen core open-source software/infrastructure globally
• Quickly identify and eliminate vulnerabilities before public release

Participation (limited alliance)

• Participation is described as limited, with roughly ~40–50 organizations/institutions selected.
• Included actors are described as large tech/security and infrastructure-heavy organizations (e.g., AWS, Apple, Cisco, CrowdStrike, Google, JP Morgan, etc.), along with mentions of foundations/projects like Linux and security vendors.

Cost / funding points

• Anthropic reportedly provides token-use funding (reported as ~$100M / ~150B KRW).
• A separate fund is mentioned (~$4M / ~6B KRW) for securely managing open-source security work, described metaphorically as a “big brother” role.

---

Key technical claim: the “breakthrough” is exposure/exploitation capability

The CEO argues that while AI-assisted vulnerability discovery existed before, Mythos/Glasswing’s differentiator is “exposure capability.” In their framing, this means:

• Moving from:
  • “Find a suspected vulnerable point”
  • to “Determine whether it’s actually attackable,” assess severity, and generate weaponized exploit steps/code (where feasible)”

They also emphasize:

• Vulnerability finding ≠ weaponization Exploitability and impact analysis are treated as a more expert-heavy stage.

• Automation changes the speed/scale of the attack path. Even if turning findings into working exploit code is still difficult, the overall pipeline compresses the timeline.

---

Why this is perceived as a threat (even though framed as defensive)

Major reasons given:

1. Attackers adopt the same capabilities Models plus operational know-how can enable larger-scale, faster attacks.

2. Human review/verification exists, but can be replicated Even if the process includes verification, attackers can likely build similar pipelines.

3. Time compression increases systemic risk Past incidents are referenced to support the idea that faster vuln→exploit cycles amplify risk.

---

Model limitations: context window still matters

• Mythos is explicitly described as having an LM context-window constraint (finite input size).
• The pipeline is not portrayed as “dump a huge codebase into the model once.”
• Instead, success depends on:
  • System design + orchestration
  • How code is sliced/inserted into context
  • How outputs are verified and fed back
• They claim security experts helped design and operate the full pipeline, including methodological steps for analysis flow and result verification.

---

Economic/operational reality: token costs and compute reasoning can be enormous

• Token usage is described as a major cost driver, including:
  • Output tokens
  • Reasoning/inference/tool-call tokens in agentic pipelines
• For large repositories or multi-repo systems, costs may reach:
  • tens of millions to hundreds of millions of KRW
• An example is cited from an AI cybersecurity challenge where token costs exceeded ~300M KRW (with sponsors covering it in that instance).
• The ROI argument is that costs can still be favorable versus hiring top experts and the time required—but it’s not easy for small budgets to replicate.

---

Critique of “just use GPT/Claude APIs” as an in-house approach

The CEO argues that while teams try building pipelines by integrating LLM APIs, in practice:

• It works best only for small/simple codebases.
• For millions to tens of millions of lines, multi-repo systems, and cross-file/cross-module interactions:
  • code analysis doesn’t decompose cleanly by file/module
  • vulnerabilities often require understanding data flow and multi-component preconditions

They describe three obstacles to replicating Glasswing-like results:

1. Technical limitation Context management and orchestrating whole-program reasoning.

2. Manpower Few people have combined skills in security system design and AI pipeline engineering.

3. Token costs Reasoning and iterative tool calls scale cost nonlinearly.

---

Theori’s position (alternative approach without Mythos)

The video’s later portion contrasts Glasswing/Mythos with Theori’s approach, claiming:

• Similar results may be achievable using public commercial models
• The key is investing in system design (“harness/orchestration/scaffolding”) rather than relying only on model size

Claimed system capabilities

Theori claims their system is optimized for large-scale vulnerability discovery and includes:

• Target selection
• Guided prompting options
• Result verification
• Automation of the end-to-end pipeline so users provide code and receive:
  • reports
  • remediation suggestions

White paper / objective analysis

• They claim the work reflects objective analysis after Mythos release.
• A white paper is published describing their findings.

---

Claimed experiment vs. Mythos and additional results

• The CEO claims Theori ran their system on the same open-source targets Mythos analyzed.
• They claim full coverage of the vulnerabilities Mythos found, including:
  • without “hinting” the model with cues like “there is a vulnerability here”
  • by directly asking it to find vulnerabilities targeting the code
• They further claim:
  • 12 additional vulnerabilities
  • detection of zero-days (as described in subtitles)

---

“AICC / AI Cyber Challenge (DARPA/DOD)” background

• They explain AICC as a U.S. DoD/DARPA-style competition focused on whether AI can find vulnerabilities in very large open-source codebases.
• A key point mentioned: for finalists, deliverables/results had to be open-sourced after the competition.
• They say Theori incorporated ideas/outputs into their product (with “Zint Code” mentioned as part of their ecosystem).

---

Strategic “what defenders should do now” (3-point plan)

The CEO concludes with a defender-oriented urgency model:

1. Discover weaknesses before attackers Use AI-based analysis tools continuously, because human review can’t keep up.

2. Prioritize and patch quickly once discovered Since vulnerability volume is huge, use classification/risk scoring plus automated patch workflows.

3. Build sustainable governance and continuous verification “One-time” or “yearly” audits won’t work; security must be integrated into the development/lifecycle pipeline through continuous monitoring, updating, and operational processes.

They argue attacker costs are trending toward near-zero due to agentic automation, so defense must match that pace.

---

Project Canopy (planned broader alliance beyond Glasswing)

They mention a follow-on/alternative initiative: Project Canopy, described as:

• A broader alliance to protect the ecosystem, not a restricted set of participants
• A “tsunami defense” style cooperative defense network

Funding strategy

• Canopy uses contributed funds to cover basic token costs for scanning public infrastructure in the public interest (explicitly not for operating profit).

Governance and partnerships

• Governance is described as potentially expanding into a broader consultative body.
• Partnerships include:
  • public sector + private firms + security companies
• Rationale: public infrastructure has high societal impact and is resource-constrained.

Inclusion intent

• It is described as more inclusive than Glasswing:
  • not limited to only Korean entrants
  • expansion to APEC/Asia and Europe is referenced

---

Main speakers / sources

• Sejun Park — CEO of Theori (primary speaker; discusses offensive security implications, technical feasibility, costs, and Theori’s approach)
• Gwan-su — introduced as a CISO (joins the discussion)
• Park Kwan-soon — introduced as Theori Seesaw (co-host / interviewer voice)