Overview of Hard Fork (The New York Times)
This episode of Hard Fork (hosts Kevin Roose and Casey Newton) covers three main items: Anthropic’s preview model Claude Mythos and its cybersecurity implications (Project Glasswing); a deep-dive conversation with New Yorker writers Ronan Farrow and Andrew Marantz about their profile “Can Sam Altman Be Trusted?”; and the light “One Good Thing” segment (NASA’s Artemis 2 and a new weather app, Acme Weather). The show includes disclosures (NYT’s litigation against OpenAI; a host’s family connection to Anthropic), a Hard Fork Live event announcement, and practical security advice for listeners.
Anthropic, Claude Mythos, and Project Glasswing
-
What Anthropic announced
- Anthropic previewed a new model, Claude Mythos (Project Glasswing), and said it is too dangerous for public release.
- Instead of general release, Anthropic is granting defensive-access to a consortium of major tech and infrastructure companies (Cisco, Broadcom, Microsoft, Apple, Amazon, etc.) to harden systems before broader exposure. OpenAI and Meta are not in the consortium.
- Anthropic is also reportedly making a large amount of compute/credits available to the consortium.
-
Why this matters (technical & practical)
- The model reportedly found zero-day vulnerabilities that humans and automated tools missed — e.g., a 27‑year‑old OpenBSD bug and an FFmpeg exploit that prior scanners didn’t catch.
- Anthropic claims autonomous chaining of exploits and advanced reasoning that make large‑scale vulnerability discovery far faster and more effective than human teams.
- Security experts (e.g., Alex Stamos) view this as a potential “reset” for cybersecurity — defenders may have a runway to patch critical systems, but there’s also the risk that models could invent new, unseen exploits.
-
Two broad future scenarios (from interviewed experts)
- Optimistic: finite set of critical vulnerabilities can be identified and patched with concentrated effort.
- Pessimistic: models can invent novel exploits continually, creating a persistent, growing security problem.
-
Legal, regulatory, and national security context
- Anthropic’s restraint appears driven by liability, public safety, and self‑interest rather than pure marketing — releasing such a tool broadly could enable rapid weaponization and legal fallout.
- The U.S. government has declared Anthropic a supply‑chain risk and banned Claude use by federal agencies, creating the weird situation that potentially critical defensive tech exists but government access is limited.
- Model development of this power remains largely unregulated in the U.S., which the hosts flag as alarming.
-
Practical advice for listeners
- Do basic cybersecurity hygiene: use a password manager (random, unique passwords), enable multi‑factor authentication (prefer authenticator apps), and promptly install software updates and patches.
- Expect more patch/update prompts across apps and devices in the coming months.
-
Notable line: Project Glasswing invoked the glasswing butterfly metaphor (transparent wings, hiding in plain sight).
Sam Altman profile: highlights from Ronan Farrow & Andrew Marantz (The New Yorker)
-
Piece title and framing: “Can Sam Altman Be Trusted?” — a long-form, heavily sourced portrait documenting patterns of behavior and disputes around Altman and OpenAI.
-
Key new revelations / reporting beats
- The reporting suggests the “outside investigation” after Altman’s 2023 firing did not produce an accessible written report; rather, the review was kept out of writing, contrary to expectations.
- The authors document deeper and longer‑running concerns (and more corroborating sources) about Altman’s repeated misstatements, selective truths, and behavior than was previously public.
- Reporting strengthens earlier claims about fundraising and relationships in the Gulf (Emirati and Saudi ties) being more extensive than previously known.
- Some former allies and board members who initially supported Altman now say they would have acted differently if they had known what they know now.
-
Range of views captured
- The piece is deliberately forensic and even‑handed: it includes people who strongly dislike Altman, those who still support him, and discussion of smear campaigns (notably amplified by Elon Musk and allies).
- The authors emphasize there is no single “smoking gun”; rather, a cumulative pattern emerges from many incidents, memos (e.g., Dario Amodei, Ilya/Sutskever), and interviews.
-
Notable quotes & characterizations reported
- A Microsoft executive: “a small but real chance he’s eventually remembered as a Bernie Madoff or Sam Bankman‑Fried–level scammer.”
- An unnamed board member: Altman is “unconstrained by truth” with “an almost sociopathic lack of concern for the consequences.”
- The profile includes lighter evidence of inconsistency (the gray-sweater anecdote) alongside serious allegations.
-
Implications for OpenAI and governance
- The story raises governance and succession questions: Can OpenAI scale into a mature public company with concentration of power around one charismatic founder?
- OpenAI’s recent moves (acquiring media outlets, announcing safety fellowships, new governance announcements) are read as efforts to shape narrative and address external scrutiny.
- The piece underscores systemic issues: the tech ecosystem currently relies heavily on private-company discretion rather than robust regulatory or oversight frameworks.
One Good Thing (lighter segment)
-
Kevin’s pick: NASA’s Artemis 2 mission
- Artemis 2 will orbit the moon and travel further from Earth than any human mission in decades (~252,756 miles away).
- Kevin praises the mission’s wonder and urges robust support for NASA.
-
Casey’s pick: Acme Weather (new weather app)
- Built by former Dark Sky team members; focused on probabilistic forecasts and community-sourced reports.
- Features: range charts showing forecast uncertainty, push notifications for lightning, sunsets, sunscreen alerts, umbrella reminders, and community-reported aurora/rainbow alerts.
- Arrives on iOS now (paid subscription) with Android planned.
Announcements & disclosures
- Hard Fork Live II: June 10 in San Francisco at the Blue Shield of California Theater; ticket sales open Friday, April 17 at nytimes.com/events.
- Disclosures on the episode: hosts mentioned NYT’s lawsuit against OpenAI/Microsoft/Perplexity and one host noted a fiancé working at Anthropic.
- Staff note: Executive producer Jen Poyant is leaving the show (farewell shout-out).
Main takeaways and recommended actions
- The Claude Mythos / Project Glasswing story is a watershed moment for cybersecurity and AI safety: defenders may get short-term benefits, but the risk landscape could fundamentally change.
- Companies, governments, and regulators need better coordination and possibly new oversight frameworks to manage AI systems that can autonomously discover and chain exploits.
- Individual actions you can take now:
- Use a password manager and unique random passwords.
- Enable multi‑factor authentication (prefer app-based authenticators).
- Keep devices and apps updated; accept security patches promptly.
- If you want the fuller context: read the New Yorker piece on Sam Altman for the full sourcing and narrative; watch the full episode on YouTube (youtube.com/hardfork) for the full interviews and tone.
Notable moments / memorable lines
- Project name symbolism: “Glasswing” — transparent wings, hiding in plain sight.
- Host confession: a pragmatic take — Anthropic’s non‑release is plausibly safety/legal self‑interest rather than marketing theater.
- Entertainment/closure: light cultural moments (gray-sweater anecdote, rainbow detector), and hosts’ genuine enthusiasm for space and clever consumer apps.
If you want a one-line summary: Anthropic’s withheld model marks a seismic cybersecurity moment that could force a massive global patch‑up — while the deep New Yorker profile raises major governance and trust questions about Sam Altman and OpenAI.