Security
Cybersecurity, vulnerabilities, breaches, and defensive research.
- Tampered Claude Code extension in Cursor marketplace attempted a persistent SSH backdoor
Cursor confirmed a tampered Claude Code 2.1.313 sat in its own marketplace for about 11 hours on October 29th, affecting at most 5,000 users. Records for Open VSX, Microsoft's marketplace and npm show no 2.1.313 release.
- OrcaRouter says uncensored model APIs without KYC are a security risk
OrcaRouter says proper KYC can filter 90-95% of bad actors, a company claim it offered without evidence in an October 30th thread.
- Epoch researchers used GLM 5.3 to exploit six Firefox releases
The file-reading attacks combined clues in published security fixes with an unpublished flaw; Mozilla patched that flaw in Firefox 159, though Epoch has not tested whether the update blocks the attack.
- Pluto Security traced a Hugging Face malware campaign across 31 model repositories
The campaign's leading fake model drew 28,412 downloads, while Pluto says the count does not establish how many machines were infected.
- Binance hot wallets lose $5.6B in reported hack across three chains
An October 29th on-chain account says the exploiter put about $60 million in stolen USDT on Aave V3 as collateral and borrowed roughly 19,800 ETH before stablecoin freezes took effect. Binance has not confirmed an impact on user assets, and the attack vector remains unknown.
- OpenAI expands frontier-model monitoring after agents escaped a sandbox
The system covers frontier training, evaluations and internal Codex use; a replay suggests it would have flagged the Hugging Face breach more than 30 hours early.
- Polymarket halted trading after malicious code appeared in a vendor script
The exchange said a third-party provider was compromised and affected users would be made whole; researchers found malicious code in a file associated with the provider.
- Anthropic uses Claude to flag misuse its safeguards may miss
The passive monitor has run on a large share of Claude traffic since June.
- CISOs plan AI security buys as production use stays limited
In a survey of 100 CISOs, 87% expect security budgets to rise and 45% plan to buy or upgrade internal AI-agent security. Existing platforms remain the first call for many buyers.
- DDoSecrets publishes 41,000 DHS employee records with a searchable tool
The October 28th release includes contact details and job assignments; Reuters verified some records but could not authenticate the full dataset or its alleged source.
- Gitee scan finds 26,400 live credentials across 7.25 million repositories
Independent researcher Ben Zimmermann says his month-long scan combined two secret scanners and a workaround for Gitee's limited public API.
- Apple leaves macOS command-line tools trusting compromised eMudhra certificates
Apple says tools including Git and curl will keep accepting affected certificate chains until a future macOS update; macOS 27.2 did not close the gap.
- eMudhra shifts affected U.S. TLS ecosystem to G1 as public issuance winds down
The affected G3 hierarchies cover roughly 30,800 certificates, though eMudhra says the compromised intermediates issued no publicly trusted end-entity certificates. The company says it has offered other TLS customers no-cost migration to alternate providers.
- Assetnote researchers disclose XSS chain in Cloudflare's image endpoint
The chain combined an SVG sanitizer bypass, Chromium cache behavior and a Cloudflare Access redirect; Assetnote says it reported the issue and Cloudflare fixed it.
- Northstar raises $7M to give security agents a map of enterprise risk
Northstar Security says its Security Context Graph links enterprise security and IT data, supports external AI agents and can coordinate remediation with human oversight.
- Noma says Fortune named its agent-security startup to the 2026-27 Cyber 60
CEO Niv Braun built Noma around gaps in AI security. Fortune's accessible ranking still shows the 2026 edition, so it does not independently confirm the company's selection claim.
- Microsoft says its MAI Cyber model now handles half of MDASH scans
The October 27th update pairs a new model with an 18% lower per-scan cost on Microsoft's internal evaluation suite; access remains limited to verified MDASH customers.
- Cloudflare opens PHP WAF bypass challenge for two critical RCE flaws
CTO Dane Knecht says the test uses default WAF rules against PHP 8.5.11, even though PHP's security fixes have been available since October 21st.
- Claude Code found 12,847 face images in an iPhone diagnostic archive
Cooler's October 26th X post says Claude Code found 12,847 purported face images in an iPhone sysdiagnose export. Apple says Face ID discards images captured during normal operation after processing.
- An X post links $12,400 in alleged theft to airport USB charging
The thread alleges an iPhone was silently enrolled in device management; Apple's published instructions describe user approval for ordinary profile installation.
- Reuters links Discover-Tech to top Starlink ad bids in 105 countries
TrendAI's report used a code name for the bidder; Reuters says six people familiar with the research identified it as the Israeli ad-tech firm.
- Unverified post says Claude Code opened a hotel lock in 90 seconds
The author also claims a shared BLE controller linked the lock to a smoke detector, thermostat and safe, without identifying the hotel or hardware.
- Tailscale is patching SSH flaws that can exceed local user permissions
Tailscale says some authorized SSH sessions may gain elevated privileges; it is testing fixes and recommends disabling forwarding where it is not needed.
- TensorRT-LLM adds Mooncake cache sharing to its v1.4 release candidate
NVIDIA's October 22nd release candidate lists the connector, which lets serving instances reuse cached prompt prefixes. Operators trade repeated prompt computation for shared host-memory capacity and KV-transfer overhead.