Anthropic has crossed a line
In a Nutshell
Anthropic's Mythos model vastly outperforms Claude Opus 4.6 on benchmarks, with its Glasswing preview exposing thousands of day-zero vulnerabilities in systems like FreeBSD and FFmpeg, shared privately with tech giants to patch before public release. While enabling hackers to craft advanced exploits and phishing, it empowers defenders more through automated vulnerability scanning, log analysis, and best practices, ultimately automating cybersecurity to overcome human weaknesses like inconsistency and phishing susceptibility. Long-term, AI levels the playing field but favors "good guys" hardening systems, marking an exciting automation era without doomsday scenarios.
These notes were generated by AI and may contain inaccuracies.
Mythos is the next model from Anthropic, standing head and shoulders above Opus 4.6 according to benchmarks. Some benchmarks jump by about 20%, going from 80% saturation to over 93%, others from 53% to 65%. This represents a step change similar to ChatGPT 3.5 to ChatGPT 4 or 4.5. Rumors indicate it's a 10 trillion parameter model, showing continued scaling. This progress means AI is not hitting a wall in net effect.
Glasswing is the Mythos preview version given early access to about 40 titans of industry like Microsoft and Oracle. It poses a strong cybersecurity risk, so Anthropic provides this access to allow big tech companies to catch up, patch vulnerabilities, and add safeguards before general availability.
AI has reached a point where it can help hackers write code and analyze code bases to find new vulnerabilities. Mythos found thousands of day zero exploits, not published, but vendors are contacted privately. Examples include a 27-year-old vulnerability in FreeBSD, a highly secure Unix operating system used on mainframes, and a 16-year-old exploit in FFmpeg, an audio-video codec driver used widely.
This is significant for information technology, cybersecurity, and national defense. Recent news reports suggest China launched a semi-automated botnet attack powered in part by Anthropic's Claude, though not personally verified.
Speaker was an IT infrastructure engineer focused on cloud automation, responsible for keeping data centers running, integrating with database, application, and cybersecurity teams.
AI is not used directly as a hacking tool but to write hacking tools, scripts, interpret telemetry from scripts to find vulnerabilities.
- Software security: Programs like Microsoft Windows, Word, Outlook.
- Network security: Cisco switches, firewalls, VLANs.
Both have different vulnerabilities and best practices.
Chief Information Security Officers (CISOs) panic and order best practices reviews, working with vendors like VMware, Microsoft, Dell, Oracle, IBM, Cisco. Fortune 500 companies outsource expertise and rely on vendor security bulletins, newsletters on vulnerabilities, mitigations, and best practices that limit attack surface. Many vulnerabilities require elevated or execution access to exploit, so best practices mitigate most issues.
New class of hackers with day zero exploits, but companies following best practices mitigate most ahead of time. Short-term, tools like Mythos provide structural advantages to attackers due to faster adaptation. Long-term, good guys (Huntress, antivirus, network vendors, Cisco, Dell, IBM, OS writers) use the same tools to secure platforms, leading to better best practices, analyzers, automated penetration testing, stateful email/packet inspection.
Vendors and internal teams use AI to analyze security logs, network logs, identify vulnerabilities. Every Fortune 500 has a perimeter of vendors using these tools to harden products.
Humans are the weakest link (OSI layer 8, PEBKAC, ID 10 T). Most expensive breaches from tricking humans via phishing. Often not executives but lawyers, doctors; smart people trap makes them easy to dupe. Lawyers split into those who trust experts vs. those who think they know better (latter easier to work with).
Attackers use AI for better phishing emails. Defenders use rules-based engines (whitelists, blacklists, vendor checks, malformed headers) plus AI to read email chains, check sender history in database, automate sanity checks server-side or client-side.
Like Subaru's twin cameras detecting lane drift or slowing cars, or full self-driving, AI automates cybersecurity behaviors humans should do, raising the standard for consistency.
Humans are inconsistent, lazy, complacent, needing annual training. Scripts run consistently every time. Previous automation was fixed; AI adds dynamism, heuristics, log reading, vendor documentation search for new bulletins/vulnerabilities.
Humans have always been clever primary attackers. AI better coders than most/all humans soon. Automates what humans can't (e.g., packet checking, email security). All AI/LLMs/robotics are new automation for enterprises: log shipping, patches, backups, now more.
Both good guys (organizations, governments) and bad guys (foreign attackers) get tools, but more energy into hardening than weakening. Use LLMs for automated security training, phishing/penetration testing, social engineering tests, reading emails/chat logs to detect manipulation.
Future has AGI in networks like Ghost in the Shell attack barriers (firewalls) with generative models writing code. Cool time in technology; still need humans in loop short-term, but long-term humans become negative EV due to slowness/mistakes.
Professional take on Mythos and Glasswing.
Keep David Shapiro in your library
Save the videos and channels worth coming back to, and find them again in one place.
