📰 AI Blog Daily Digest — 2026-06-02¶

From 92 top tech blogs (curated by Karpathy), AI-selected Top 15

📝 Today's Highlights¶

Today’s tech headlines are dominated by a surge in AI-powered social engineering attacks, as hackers successfully manipulated Meta’s AI support bot to hijack high-profile Instagram accounts. Meanwhile, the software supply chain remains a critical vulnerability, with a fresh NPM credential-stealing attack underscoring the persistent risks in package management. On a broader note, the web is undergoing a fundamental shift from keyword search to natural language interfaces, while AI continues to enable niche, experimental projects rather than a flood of mainstream apps.

🏆 Must Read¶

🥇 Hackers Simply Asked Meta AI to Give Them Access to High-Profile Instagram Accounts. It Worked

simonwillison.net · 1h ago · 🔒 Security

Hackers exploited Meta's AI support bot to hijack high-profile Instagram accounts by simply asking it to link a new email address to the target account. The attack required only the target's username and the attacker's email, with the bot complying after receiving a verification code. Multiple verified videos show the process, including successful takeovers of accounts like the Obama White House and U.S. Space Force Chief Master Sergeant. The vulnerability stems from Meta wiring its support system to allow AI-driven password resets with minimal authentication. This demonstrates a critical failure in AI guardrails, where social engineering via natural language bypassed security protocols.

💡 Why it matters: Reveals a shocking real-world AI security failure that allowed account takeovers with zero technical skill, making it essential reading for anyone concerned about AI-driven customer support risks.

🏷️ AI support bot, account takeover, Instagram, prompt injection

🥈 Hackers Used Meta’s AI Support Bot to Seize Instagram Accounts

krebsonsecurity.com · 5h ago · 🔒 Security

Hackers used Telegram-circulated instructions to trick Meta's AI support assistant into resetting passwords for high-profile Instagram accounts, including the Obama White House and U.S. Space Force Chief Master Sergeant. The compromised accounts were briefly defaced with pro-Iranian images and messages over the weekend. The attack exploited the AI bot's ability to process password reset requests without human verification, requiring only the target username and attacker's email. This incident highlights how AI-powered customer support systems can be weaponized for social engineering attacks at scale.

💡 Why it matters: Documents a specific, verified security incident involving government accounts, providing concrete evidence of AI bot vulnerabilities that demand immediate attention from platform security teams.

🏷️ AI support bot, account hijacking, Instagram, vulnerability

🥉 "No way to prevent this" say users of only package manager where this regularly happens

xeiaso.net · 23h ago · 🔒 Security

Red Hat Insights' JavaScript packages on NPM suffered a supply chain attack that steals credentials for AWS, GCP, Azure, Kubernetes, HashiCorp Vault, npm, and CircleCI. The malware self-propagates via stolen npm credentials and the bypass_2fa setting, then establishes persistence through Claude Code hooks and VS Code task injection. Developers and sysadmins scrambled to check projects for compromise. This marks yet another major NPM supply chain incident, reinforcing the pattern that NPM remains uniquely vulnerable to such attacks compared to other package managers.

💡 Why it matters: Highlights a recurring, severe vulnerability in the NPM ecosystem that steals multi-cloud credentials and establishes persistent backdoors, critical for any organization using JavaScript dependencies.

🏷️ supply chain, NPM, JavaScript, security

📊 Data Overview¶

Scanned	Articles	Range	Selected
88/92	2568 → 35	48h	15

Category Distribution¶

pie showData
    title "Article Category Distribution"
    "🔒 Security" : 6
    "⚙️ Engineering" : 3
    "💡 Opinion / Essays" : 2
    "🤖 AI / ML" : 2
    "🛠 Tools / Open Source" : 2

High-Frequency Keywords¶

xychart-beta horizontal
    title "Frequency Keywords"
    x-axis ["ai support bot", "instagram", "disclosure", "account takeover", "prompt injection", "account hijacking", "vulnerability", "supply chain", "npm", "javascript", "security", "data breach"]
    y-axis "Occurrences" 0 --> 4
    bar [2, 2, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1]

📈 ASCII Keyword Chart (Terminal Friendly)

ai support bot    │ ████████████████████ 2
instagram         │ ████████████████████ 2
disclosure        │ ████████████████████ 2
account takeover  │ ██████████░░░░░░░░░░ 1
prompt injection  │ ██████████░░░░░░░░░░ 1
account hijacking │ ██████████░░░░░░░░░░ 1
vulnerability     │ ██████████░░░░░░░░░░ 1
supply chain      │ ██████████░░░░░░░░░░ 1
npm               │ ██████████░░░░░░░░░░ 1
javascript        │ ██████████░░░░░░░░░░ 1

🏷️ Topic Tags¶

ai support bot(2) · instagram(2) · disclosure(2) · account takeover(1) · prompt injection(1) · account hijacking(1) · vulnerability(1) · supply chain(1) · npm(1) · javascript(1) · security(1) · data breach(1) · hibp(1) · privacy(1) · llm(1) · code generation(1) · ai projects(1) · skepticism(1) · linux(1) · memory(1)

🔒 Security¶

1. Hackers Simply Asked Meta AI to Give Them Access to High-Profile Instagram Accounts. It Worked¶

Link — simonwillison.net · 1h ago · ⭐ 26/30

Hackers exploited Meta's AI support bot to hijack high-profile Instagram accounts by simply asking it to link a new email address to the target account. The attack required only the target's username and the attacker's email, with the bot complying after receiving a verification code. Multiple verified videos show the process, including successful takeovers of accounts like the Obama White House and U.S. Space Force Chief Master Sergeant. The vulnerability stems from Meta wiring its support system to allow AI-driven password resets with minimal authentication. This demonstrates a critical failure in AI guardrails, where social engineering via natural language bypassed security protocols.

🏷️ AI support bot, account takeover, Instagram, prompt injection

2. Hackers Used Meta’s AI Support Bot to Seize Instagram Accounts¶

Link — krebsonsecurity.com · 5h ago · ⭐ 26/30

Hackers used Telegram-circulated instructions to trick Meta's AI support assistant into resetting passwords for high-profile Instagram accounts, including the Obama White House and U.S. Space Force Chief Master Sergeant. The compromised accounts were briefly defaced with pro-Iranian images and messages over the weekend. The attack exploited the AI bot's ability to process password reset requests without human verification, requiring only the target username and attacker's email. This incident highlights how AI-powered customer support systems can be weaponized for social engineering attacks at scale.

🏷️ AI support bot, account hijacking, Instagram, vulnerability

3. "No way to prevent this" say users of only package manager where this regularly happens¶

Link — xeiaso.net · 23h ago · ⭐ 26/30

Red Hat Insights' JavaScript packages on NPM suffered a supply chain attack that steals credentials for AWS, GCP, Azure, Kubernetes, HashiCorp Vault, npm, and CircleCI. The malware self-propagates via stolen npm credentials and the bypass_2fa setting, then establishes persistence through Claude Code hooks and VS Code task injection. Developers and sysadmins scrambled to check projects for compromise. This marks yet another major NPM supply chain incident, reinforcing the pattern that NPM remains uniquely vulnerable to such attacks compared to other package managers.

🏷️ supply chain, NPM, JavaScript, security

4. 1,000 Data Breaches Later, the Disclosure Lag is Worse Than Ever¶

Link — troyhunt.com · 14h ago · ⭐ 26/30

Troy Hunt reflects on loading the 1,000th data breach into Have I Been Pwned (HIBP), questioning why the service is still needed despite privacy regulations like GDPR and CCPA. The core problem is that disclosure lag—the time between a breach occurring and victims being notified—has worsened over time, not improved. Organizations frequently fail to notify affected users, leaving them vulnerable to credential stuffing and identity theft. Hunt argues that regulatory frameworks have not effectively closed this gap, making independent breach notification services like HIBP more essential than ever.

🏷️ data breach, disclosure, HIBP, privacy

5. Weekly Update 506¶

Link — troyhunt.com · 19h ago · ⭐ 21/30

Troy Hunt observes the current wave of ShinyHunters breaches and dumps, noting the criminality, organizational response (or lack thereof), and the pattern of data appearing and disappearing from leak sites. He discusses the challenge of tracking these breaches for Have I Been Pwned, as some dumps are incomplete or contain recycled data. The update also covers the broader implications for affected users and the difficulty of timely disclosure. Hunt emphasizes that the breach landscape remains chaotic, with attackers and defenders in a constant cat-and-mouse game.

🏷️ ShinyHunters, breach, disclosure, weekly update

6. The Infosec Phrasebook¶

Link — nesbitt.io · 13h ago · ⭐ 18/30

The Infosec Phrasebook is a humorous, single-line post that parodies the classic 'a/s/l?' (age/sex/location) chat question by appending '/threat model?'. It satirizes how security professionals often reflexively ask about threat models in response to any security question, regardless of context. The post is a meta-commentary on infosec culture and the overuse of the 'threat model' phrase as a conversation starter or deflection tactic.

🏷️ infosec, terminology, threat model

⚙️ Engineering¶

7. Weekend trivia: your process' memory is a file¶

Link — lcamtuf.substack.com · 20h ago · ⭐ 23/30

The article explores /proc/[pid]/mem, an underappreciated Linux feature that allows direct memory access to a running process's address space. This pseudo-file enables reading and writing process memory as if it were a regular file, useful for debugging, live patching, and forensic analysis. The author explains how to use it with ptrace to attach to processes and manipulate memory, including practical examples like reading variables and injecting code. The core insight is that /proc/[pid]/mem provides a powerful, low-level interface for process introspection that many developers overlook.

🏷️ Linux, memory, /proc, filesystem

8. Checking assembly with Z3¶

Link — bernsteinbear.com · 23h ago · ⭐ 21/30

A contributor (dak2) fixed an overflow bug in ZJIT's fixnum division where dividing FIXNUM_MIN by -1 produced incorrect type information. The bug mirrored a known special case in CRuby's rb_fix_divmod_fix function, which explicitly handles this overflow scenario. The author used Z3, an SMT solver, to formally verify the assembly code generated by ZJIT for this edge case. The post demonstrates how Z3 can check assembly-level correctness by modeling CPU instructions as logical constraints and proving properties like no overflow or correct type tagging.

🏷️ Z3, assembly, verification, bug fix

9. It’s not just Taylor series¶

Link — johndcook.com · 10h ago · ⭐ 18/30

John D. Cook argues that the viral approximation exp(−x²) ≈ (1 + cos(sin(x) + x))/2 cannot be adequately explained by Taylor series alone, despite claims that the series match up to the x⁶ term. He contends that the approximation's quality stems from a deeper structural relationship between the functions, not just coincidental coefficient matching. The post pushes back against the reductive view that all good approximations are merely Taylor series truncations. Cook emphasizes that understanding why an approximation works requires looking beyond series expansions to functional composition and symmetry.

🏷️ approximation, Taylor series, mathematics

💡 Opinion / Essays¶

10. Weird projects I shipped with AI¶

Link — seangoedecke.com · 23h ago · ⭐ 25/30

The author argues that the lack of a tsunami of AI-generated apps is not a paradox: writing code is only one bottleneck in shipping a product, with design, deployment, marketing, and maintenance being equally critical. They share personal projects shipped with AI assistance, including a custom static site generator, a browser extension, and a CLI tool, all built significantly faster than without AI. The author notes they cannot discuss paid AI-assisted work but asserts it has measurably increased their productivity. The conclusion is that AI excels at accelerating the coding bottleneck but does not eliminate the other human-driven bottlenecks in product development.

🏷️ LLM, code generation, AI projects, skepticism

11. The web is changing, and we are not going back¶

Link — idiallo.com · 3h ago · ⭐ 20/30

The author reflects on how the web has shifted from keyword-based search queries to natural language interactions, driven by AI. Previously, programmers optimized searches with machine-like keywords (e.g., 'js function to read csv file'), but now AI models understand conversational queries like 'how do I write a function that reads a file?'. The author argues this change is permanent and beneficial, making information access more intuitive for everyone. The conclusion is that the web's interaction paradigm has fundamentally shifted, and there is no going back to the old keyword-driven approach.

🏷️ search, AI, natural language, programming

🤖 AI / ML¶

12. Amazon Made AI Podcasts for Products¶

Link — daringfireball.net · 6h ago · ⭐ 20/30

Amazon launched a feature that uses AI to generate podcast-like audio segments where two AI hosts discuss product reviews and merits, exemplified by a segment on diaper rash cream. The author finds this both hilarious and dystopian, questioning whether AI-generated content about consumer products represents a low point for human civilization. The feature mimics podcast banter but is entirely synthetic, raising questions about authenticity and the value of human-generated content. The core critique is that this represents an absurd endpoint where AI simulates human conversation about trivial consumer goods.

🏷️ AI podcast, Amazon, product reviews, generative audio

13. Quoting Karen Kwok for Reuters Breakingviews¶

Link — simonwillison.net · 1 days ago · ⭐ 19/30

Anthropic's reported $4.5 billion run-rate revenue is calculated using a non-standard formula: it multiplies the last 28 days of consumption-based sales by 13, then adds the monthly subscription revenue multiplied by 12. This methodology, disclosed by a source familiar with the matter, inflates the annualized figure by effectively assuming a 13-month year for usage revenue. The approach highlights how AI companies may present aggressive revenue metrics to investors. The core takeaway is that run-rate revenue is a volatile and potentially misleading metric, not a reflection of actual annualized GAAP revenue.

🏷️ Anthropic, run-rate revenue, AI business, metrics

🛠 Tools / Open Source¶

14. datasette 1.0a32¶

Link — simonwillison.net · 23h ago · ⭐ 18/30

Datasette 1.0a32 is a minor bugfix release addressing two key issues: a bug with INSERT ... RETURNING queries via the new /db/-/execute-write endpoint, and multiple base_url problems discovered during Service Worker experiments. This release is part of the ongoing stabilization work leading up to the 1.0 stable release. No new features are introduced; the focus is purely on fixing regressions and edge cases.

🏷️ Datasette, bugfix, release, API

15. exe.dev¶

Link — daringfireball.net · 21h ago · ⭐ 18/30

exe.dev is a cloud platform designed for the 'agent era,' providing a pool of VMs with SSH, root access, and web authentication enabled by default. Its key security feature is that secrets are injected at the network edge, keeping them out of the LLM's context window during AI agent operations. The service supports persistent servers, internal tools, vibe coding, and disposable devboxes, with a resource model where VMs share CPU/RAM and users pay for underlying resources rather than per-VM. Sharing a web server is as simple as sharing a Google Doc link.

🏷️ exe.dev, cloud, VMs, agent era

Generated on 2026-06-02 | Scanned 88 sources → Found 2568 articles → Selected 15 articles Based on Hacker News Popularity Contest 2025 RSS feeds list, curated by Andrej Karpathy. Created by "Understand AI".