CircleCI MCP Server Flaw Hits Maximum CVSS 10.0

Critical Unauthenticated RCE in CircleCI MCP Server Earns CVSS 10.0

A critical remote code execution vulnerability in CircleCI’s MCP server has been disclosed, carrying a maximum CVSS 3.1 score of 10.0. The flaw allows a fully unauthenticated attacker to execute code in the context of the organization’s CI environment, including access to secrets and identity. The issue stems from a safeguard that checks Host and Origin headers; an attacker can send a request with Host: localhost and no Origin, which passes both checks because localhost is allowed and a missing Origin is explicitly permitted. By default, the server listens on all interfaces, making it network-reachable.

CircleCI shipped a fix in version 0.19.2 and published a public advisory (GHSA-xv5j-cwgj-22r4). The timeline shows the report on 30 July 2026, fix on 6 August, and public advisory on 10 August. The source notes this is a symptom of a broader MCP ecosystem problem where dangerous functionality is a feature, not a bug, and some vendors have declined to issue advisories. Security teams are advised to correlate MCP access logs, rotate secrets if exposed, place MCP servers behind a hardened access layer, and treat pipeline execution as privileged code execution.

Unauthenticated RCE in CircleCI MCP Server Explained →

NYC DOB NOW Permit Portal IDOR Exposed All Building Documents

New York City confirmed on April 15, 2026, that a vulnerability in its DOB NOW building permit portal has been fully resolved, and the researcher who found it is now listed on the city’s page of responsible security researchers. The flaw was a classic Insecure Direct Object Reference (IDOR): the document list API returned metadata for every document associated with a filing, including internal server URLs, regardless of authorization. The download endpoint only verified authentication, not permission, so any authenticated user could download any document by substituting the URL parameter.

The researcher discovered the issue while investigating the inspection process for a major electronic music venue in Brooklyn. The vulnerability potentially affected documents across the entire system, including architectural plans, sketches, and internal filing documents for buildings across the city. The city resolved the issue on April 15, 2026, after the researcher reported it through the Vulnerability Disclosure Program with steps to reproduce and recommendations for fixing backend permission checks.

Hacking the New York City Building Permit Portal | Michael Cummings →

Three Studies Show Passkey Bypass Attacks on Windows

Three independent studies published in August 2026 demonstrated post-compromise attack chains that bypass passkey-based phishing-resistant authentication on Windows. SpecterOps, Palo Alto Networks Unit 42, and researcher Dirk-jan Mollema each identified a distinct vector. All three attacks require malware already present on the endpoint and do not require privilege escalation. None breaks FIDO2/WebAuthn cryptography; all exploit the assumption that the endpoint device is trusted.

SpecterOps demonstrated a Windows-to-Entra ID chain allowing impersonation of privileged users, exploiting that the Windows Event Logging Service stores previous YubiKey signatures in cleartext (CVE-2026-34348, CVSS 6.5). Microsoft shipped mitigations in July 2026 updates. Unit 42 identified three variants against Google Password Manager in Chrome, including Golden Pass-ta-key which targets the Security Domain Secret (SDS), a 32-byte master key that remains exposed in process memory. Dirk-jan Mollema showed a low-privilege process can use Windows Hello for Business keys without a new prompt, and that the Entra ID WebAuthn challenge is valid for 5 minutes and not bound to session, user, or tenant, allowing challenge relay. Recommended actions include patching, validating the UV flag, monitoring for anomalous device registrations, and preferring hardware-backed passkeys for privileged accounts.

Passkey Bypass: Three Attacks Demolish Phishing-Resistant… | DeafNews →

French Cyber Spies Used GitHub Malware to Hack EncroChat

French cyber spies hacked the EncroChat encrypted phone network in 2020 using malware copied from GitHub, according to an investigation by Computer Weekly and the Sussex Centre for Law and Technology. The malware relied on the Bad Binder bug (CVE-2019-2215), an Android exploit left unpatched in 2.5 billion phones for two years. The French government hacking team STNCJ downloaded Frida, an open source Android toolkit, from GitHub to build the implant, which was described as poorly written and lacking elementary countermeasures. On 1 April 2020, STNCJ injected implant code into the EncroChat update system, infecting 32,014 devices over two months.

The operation led to more than 6,500 arrests, the seizure of 270 tons of drugs, and cash finds worth nearly €1bn. Czech spyware company Invasys rehacked an infected phone and found digital fingerprints matching French evidence, proving how messages were copied. Invasys also found that the 71 EncroChat server virtual machine images given to them by the NCA had been tampered with, appearing to have been processed in gross violation of forensic principles. The UK’s Court of Appeal has ruled that messages were taken from storage, making the French data admissible as equipment interference. The disclosure is likely to restart a critical case into the legality of police tactics in Britain’s Investigatory Powers Tribunal.

Revealed: Cyber spies used malware from GitHub to hack EncroChat cryptophone network | Computer Weekly →

French Tax Agency Breach Exposes Data on 678,000 People

The French tax agency, DGFiP, stated that an attacker took data belonging to 678,000 individuals and businesses. The intrusions occurred in June and July, and the agency did not establish that data had left its systems until the attacker said so in August. For individuals, the attacker accessed reference tax income, family quotient, and withholding tax rate; for companies, the registered name and SIREN number. The attacker also consulted cadastral records covering property addresses and floor areas. The attack did not exploit a software flaw but used stolen identifiers of a DGFiP employee and an authorised third party, and the attacker described bypassing multi-factor authentication.

On 12 and 13 August, someone using the alias ZeroBytes claimed the access on a cybercrime forum and offered a database for sale. The government’s figure is 678,000 individuals and businesses, though the attacker claimed to have pulled 252,149 records covering more than two million people. Prime Minister Sébastien Lecornu chaired an interministerial crisis cell and asked ANSSI to audit the incident. The Paris prosecutor has opened an investigation covering fraudulent extraction of data and criminal conspiracy. DGFiP has referred the matter to CNIL and will lodge a criminal complaint. The exposed combination of tax income, withholding rate, and cadastral data describes household earnings and property details, which could feed scams and identity fraud.

France’s tax agency lost data on 678,000 people to a stolen login →