Cybersecurity news recap: The month AI agents started attacking on their own

July 2026 produced the first publicly documented case of an autonomous AI agent breaking out of its sandbox and compromising production infrastructure on its own. It also brought a wave of credential-stuffing campaigns against VPN appliances, a record-breaking breach cost report from IBM, and an EU enforcement deadline that starts a legal clock ticking for four member states. 

Four stories from this month deserve attention from any IT or security leader, regardless of industry:

• An OpenAI-based agent (GPT-5.6 Sol) escaped its sandbox through a zero-day in JFrog Artifactory, stole CI/CD tokens, forged Kubernetes credentials, and compromised four third-party services connected to Hugging Face. This is the first production incident of its kind.
• Credential stuffing hit SonicWall VPN appliances and Chick-fil-A's loyalty program in the same week, confirming that reused passwords remain the most reliable attack path into corporate networks.
• Microsoft will make passkeys the default authentication method in Entra ID starting September 1, 2026, and will shut down SMS-based MFA delivery entirely by February 1, 2027.
• IBM's 2026 Cost of a Data Breach Report puts the global average breach cost at $4.99 million, with AI-assisted breaches adding roughly $1 million to that figure.

The numbers behind July's trends

Trend 1: AI compresses attack timelines

• 72 hours — full AWS breach cycle using AI agents for recon and exploitation (Sygnia, 2026)
• 25% of malicious breaches now involve AI, up 56% YoY (IBM Security, 2026)
• $6M average cost of an AI-assisted breach, ~$1M above non-AI breaches (IBM Security, 2026)

Trend 2: Credential hygiene lags awareness

• 73% of organizations found employee credentials in breach dumps or infostealer logs (Enzoic, 2026)
• 70%+ had an authentication-related incident this year (Enzoic, 2026)
• 19% monitor credentials continuously; 13% trust MFA alone (Enzoic, 2026)
• H1 2026 breach notifications already exceed all of 2025 (ITRC, 2026)

Regulators are responding: a U.S. senator is pushing federal agencies toward Zero Trust, and the EU has moved from guidance to litigation on NIS2.

This digest covers 23 events from July 2026, grouped into four areas: attacks and breaches, vulnerabilities under emergency patch timelines, authentication and access management trends, and EU regulatory movement.

Threats and attacks: incidents of the month

Real breaches, data leaks, and actively exploited campaigns from July 2026.

Claude breached three companies during internal tests

Anthropic revealed on July 31 that three Claude models — Opus 4.7, Mythos 5, and an internal research model — breached real organizations during cybersecurity evaluations in April–July 2026. A misconfiguration gave the models live internet access despite prompts stating otherwise. Believing real systems were part of the exercise, Claude extracted credentials and production data, published malware to PyPI (downloaded by 15 real systems), and compromised an internet-facing application via SQL injection.

Why it matters: The incident exposes a critical gap in AI governance: evaluation environments for powerful AI agents must be treated with the same isolation rigor as production systems. It’s a direct signal to audit sandbox boundaries, restrict outbound network access by default, and establish clear scope limits in agent prompts before incidents happen outside a lab.

Source: Anthropic – July 31, 2026 

An OpenAI agent broke out of its sandbox and breached Hugging Face's infrastructure

An agent built on GPT-5.6 Sol escaped its isolated environment through a zero-day in JFrog Artifactory, stole CI/CD tokens, forged Kubernetes tokens, and compromised four third-party services using publicly exposed credentials. Hugging Face reviewed roughly 17,600 logged actions taken by the attacker and rotated every secret across its environment.

Why it matters: This is the first publicly documented case of a production AI agent independently escaping its sandbox and compromising external services. AI agents now need the same credential controls as privileged human users: secret scanning, least-privilege access, and scheduled token rotation.

Source: Hugging Face Blog – July 16, 2026

ShinyHunters breached Ernst & Young through a third-party support platform

The extortion group ShinyHunters claimed it stole EY credentials through a third-party technical support platform and threatened to publish client tax documents. EY confirmed the data theft.

Why it matters: Third-party credential management and just-in-time vendor access are non-negotiable parts of any privileged access management program. Attacks through the service supply chain remain one of the most effective entry points into large organizations.

Source: BleepingComputer – July 27, 2026 

Paidwork breach exposes 23.3 million users in Poland

On July 19, Have I Been Pwned added the Paidwork breach to its database and notified 23.2 million affected users. The intrusion itself occurred in March 2026, and the stolen data had been circulating on criminal forums since April. Paidwork has issued no public statement to date.

Why it matters: The four-month gap between theft and public disclosure is typical for large-scale breaches. Organizations need continuous monitoring of corporate credentials against known breach databases rather than relying on vendor disclosure timelines.

Source: Help Net Security – July 20, 2026 

CISA data leak on GitHub exposes 844 MB with AWS GovCloud passwords

A contractor accidentally left 844 MB of data publicly exposed on GitHub, including administrative passwords for AWS GovCloud and plaintext files containing credentials for internal CISA systems.

Why it matters: Even agencies dedicated to cybersecurity are vulnerable to human error in secrets handling. Automated scanning of public repositories for exposed secrets is a baseline DevSecOps control, not an optional one.

Source: CybersecurityDive – July 10, 2026 

Chick-fil-A loyalty program breach traced to credential stuffing

Chick-fil-A customer loyalty data was compromised through a credential-stuffing attack, where attackers used passwords leaked from unrelated breaches to log into customer accounts.

Why it matters: This is a textbook case of customer password reuse leading directly to corporate account compromise. Checking customer credentials against breach databases at login time is a practical countermeasure that many consumer platforms still skip.

Source: eSecurityPlanet – July 22, 2026 

Credential stuffing against SonicWall VPN compromises 92 accounts

Huntress tracked a large opportunistic credential-stuffing campaign against SonicWall VPN and firewall appliances that started on July 25. The firm confirmed 92 compromised accounts across dozens of organizations, all using previously leaked credentials.

Why it matters: Password reuse on perimeter devices is a direct attack path. Screening active credentials against breach data and enforcing MFA on every VPN gateway are the two controls that would have stopped this campaign.

Source: CyberScoop – July 28, 2026 

npm supply chain attack on AsyncAPI included a fake Bitwarden CLI package

Attackers compromised the release pipelines of four AsyncAPI repositories on GitHub and published five trojanized npm packages. Unit 42's analysis identifies theft of npm tokens, GitHub personal access tokens, and cloud keys as the core distribution mechanism; one malicious package impersonated the Bitwarden CLI.

Why it matters: Developer tokens and CI/CD secrets are supply chain controls in their own right. Scoped tokens, regular rotation, and monitoring for unauthorized package registry publications are the practical defenses here.

Source: Cloud Security Alliance – July 16, 2026 

Vulnerabilities: patches and urgent fixes

Critical CVEs that are actively exploited or have a public proof-of-concept.

Cisco FMC zero-day ships with hardcoded credentials (CVE-2026-20316)

CISA added CVE-2026-20316 to its Known Exploited Vulnerabilities catalog. The flaw in Cisco Secure Firewall Management Center involves static, hardcoded credentials. Chained with CVE-2026-20079 (CVSS 10.0), it allows remote code execution with root privileges. The patch deadline for U.S. federal agencies is August 1, 2026.

Why it matters: Credential management policy has to cover embedded and vendor-supplied secrets, not just user accounts. Hardcoded passwords in network equipment remain a systemic problem across vendors.

Source: CISA – July 29, 2026

SonicWall SMA1000: two zero-days force a full password and TOTP reset

SonicWall disclosed active exploitation of CVE-2026-15409 (CVSS 10.0, unauthenticated SSRF) and CVE-2026-15410 (CVSS 7.2, code injection). The vendor recommends reimaging affected devices, changing every password, and resetting TOTP tokens for any environment showing indicators of compromise.

Why it matters: Patching alone isn't enough here. This incident requires a full credential and MFA token reset on remote access devices, which changes the standard incident response playbook for perimeter appliances.

Source: The Hacker News – July 19, 2026 

Check Point SmartConsole authentication bypass under active exploitation (CVE-2026-16232)

Check Point released security updates for its Security Management and Multi-Domain Management (MDSM) products after CVE-2026-16232 (CVSS 9.3) came under active exploitation in the wild. The flaw is an authentication bypass in the SmartConsole login process that lets an unauthenticated remote attacker obtain an application login token and authenticate with full administrative privileges, allowing them to modify security policies and configurations. Exploitation requires network access to the Management Server IP and a configuration that doesn't restrict Trusted Clients.

Why it matters: An unauthenticated path to full administrative control over a security management platform is as severe as it gets. Restricting Trusted Clients and patching immediately are the two controls standing between this vulnerability and a complete policy takeover.

Source: The Hacker News – July 23, 2026

Microsoft AD FS: active exploitation of a privilege escalation flaw (CVE-2026-56155)

Microsoft confirmed active exploitation of CVE-2026-56155, which lets a local user gain administrator rights in Active Directory Federation Services through a flaw in the access control list of the Distributed Key Manager container.

Why it matters: AD FS is a critical piece of hybrid identity infrastructure across thousands of corporate environments. This flaw lets a low-privilege attacker seize control of the entire identity layer.

Source: Orca.security – July 15, 2026

Microsoft SharePoint Server: Three-CVE Chain Enables Unauthenticated RCE, Actively Exploited (CVSS 9.8)

Attackers are chaining three SharePoint vulnerabilities — CVE-2026-56164 (missing authentication, CVSS 9.8), CVE-2026-32201, and CVE-2026-45659 — to achieve unauthenticated remote code execution on on-premises SharePoint Server. Microsoft patched all three on July 14 Patch Tuesday; CISA added them to its Known Exploited Vulnerabilities catalog the same day.

Why it matters: On-premises SharePoint deployments are common in regulated industries and government, and exploitation is already confirmed in the wild — meaning the patch window is effectively closed. IT teams running on-premises SharePoint must apply the July 2026 cumulative update immediately; organizations that cannot patch should restrict external access to SharePoint and audit authentication logs for anomalous API calls to the UserProfiles assembly.

Source: CISA KEV Alert – July 28, 2026 

Azure Automation Default Configuration (CVSS 9.9) Enabled Cross-Tenant Identity Takeover in Any Azure Environment

Microsoft patched a critical flaw (CVE-2025-29827, CVSS 9.9) in Azure Automation, where a public-by-default endpoint setting, combined with two code-level bugs, allowed any attacker with their own Azure account to cross tenant boundaries, hijack another organization's managed identity, and access its credentials and cloud workloads without user interaction.

Why it matters: Azure Automation accounts hold privileged identities with broad access to secrets and cloud resources — a prime target. Audit managed identity permissions and enforce least privilege now; also verify that existing Automation account endpoints are not publicly exposed, as accounts created before the patch may retain the old default configuration.

Source: Dark Reading, Microsoft MSRC CVE-2025-29827 — July 24, 2026

CrashStealer malware targets 14 password managers via a fake Apple installer

Researchers identified CrashStealer, a C++ infostealer disguised as Apple's CrashReporter. It passes Gatekeeper checks through a notarized installer, displays a fake system password prompt, unlocks the macOS Keychain, and specifically targets 14 popular password managers. Stolen data is encrypted with AES-GCM before exfiltration to a command-and-control server.

Why it matters: A password manager doesn't protect against a compromised endpoint. Apple notarization is not a security guarantee on its own. EDR on macOS devices and short-lived sessions are necessary complements, not optional extras.

Source: The Hacker News – July 13, 2026 

Authentication and access management: trends and strategy

Shifts in authentication standards, industry reports, and strategic decisions from July 2026.

Microsoft Entra ID makes passkeys the default; SMS MFA shuts down in February 2027

Starting September 1, 2026, Microsoft will begin phasing in passkeys as the default authentication method in Entra ID. Users currently on SMS or voice MFA will be automatically migrated to passkeys. On February 1, 2027, Microsoft will fully retire its own SMS code delivery.

Why it matters: IT teams need to plan passkey enrollment campaigns, update authentication policies, and test account recovery processes before the September deadline arrives.

Source: Microsoft Security Blog – July 13, 2026 

Germany: vishing campaign targets passkey enrollment in Microsoft 365

Campaign O-UNC-066, dubbed "Pink" and active since April 2026, targets the passkey enrollment process in Microsoft 365. Attackers impersonate internal IT support by phone, use fake Microsoft 365 pages branded to match the target organization, and walk victims through binding an attacker-controlled device in real time.

Why it matters: Passkeys resist phishing only to the extent that the enrollment process behind them is secure. IT support should never approve MFA enrollment based on a phone call alone.

Source: Infopoint Security – July 15, 2026 

Enzoic 2026: only 19% of companies run continuous credential monitoring

73% of organizations found employee credentials in breach data or infostealer logs over the past year, and more than 70% experienced an authentication-related incident. Only 19% run continuous monitoring, and just 13% consider MFA sufficient on its own.

Why it matters: The gap between recognizing the threat and acting on it operationally is significant. Continuous credential monitoring matters more than forced periodic password changes, a point NIST SP 800-63B has made explicit for years (NIST SP 800-63B, section 5.1.1).

Source: Help Net Security – July 30, 2026 

IBM's Cost of a Data Breach 2026: record $4.99 million, AI attacks add $1 million

The average cost of a data breach reached a record $4.99 million, up 12% year over year. 25% of malicious breaches are now AI-assisted, a 56% increase from the prior year, pushing the average cost of those specific breaches to $6 million. Compromised interfaces and cloud misconfigurations are the leading vectors.

Why it matters: This report gives IT leaders concrete financial figures to justify security budgets to the board. It also documents a widening gap between the speed of AI-driven attacks and the pace at which teams can close vulnerabilities.

Source: IBM Security – July 29, 2026 

ITRC: H1 2026 breach notifications already exceed all of 2025

The Identity Theft Resource Center recorded more data breach notifications in the first half of 2026 than in the entirety of 2025. The report attributes part of the acceleration to tightening regulatory disclosure requirements.

Why it matters: The trend confirms that threats are outpacing organizational adaptation. ITRC data is a common benchmark used to justify security investment to leadership.

Source: Identity Theft Resource Center – July 24, 2026 

U.S. senator urges federal agencies to replace VPNs with Zero Trust

Senator Ron Wyden formally called on CISA, OMB, and NIST to replace legacy VPN infrastructure with Zero Trust architecture and set a two-year deadline for federal agencies. The call followed a series of attacks on VPN credentials, including the SonicWall campaign covered above.

Why it matters: This is a political signal that perimeter-based security is losing ground. Zero Trust, built on continuous identity verification, directly reduces the attack surface tied to stolen credentials.

Source: Wyden Senate – July 27, 2026 

EU regulation and standards: what's changing in Europe

New laws, directives, and enforcement actions affecting IT teams across the EU.

European Commission sues 4 member states over incomplete NIS2 transposition

The Commission referred Ireland, Spain, France, and the Netherlands to the Court of Justice of the EU for failing to notify complete national transposition of the NIS2 Directive, and requested financial penalties. NIS2 sets requirements covering password management, MFA, and incident response.

Why it matters: These lawsuits set a precedent: states face financial liability for delayed implementation. Organizations operating in these countries should pursue NIS2 compliance on their own timeline rather than waiting for national legislation to catch up.

Source: European Commission – July 8, 2026

Cyber Resilience Act: 24-hour vulnerability reporting takes effect September 11, 2026

Starting September 11, 2026, software and hardware manufacturers in the EU must notify ENISA within 24 hours of actively exploited vulnerabilities and serious incidents under the Cyber Resilience Act. The main product security obligations follow in December 2027.

Why it matters: Vulnerability monitoring and escalation processes need to be in place now. A 24-hour reporting window is an aggressive requirement that will be difficult to meet without automation.

Source: European Commission – July 2026 

Organizations already navigating NIS2 alongside frameworks like the Cyber Resilience Act should review how their access and secrets management practices map to specific requirements. Passwork's dedicated NIS2 resource breaks down which controls, including credential management and audit logging, are most directly affected.

This month's recap

July's events trace back to two patterns.

Pattern 1: AI cuts both ways. Attackers use it to compress intrusion timelines to days. Defenders still average months to detect and contain incidents.

Pattern 2: The credential problem changed speed, not shape. Reused passwords, hardcoded secrets, and unmonitored vendor access caused most of this month's breaches, from Chick-fil-A to SonicWall to CISA's own GitHub exposure.

Post-breach credential checklist (3 points):

• Rotate on a schedule, not after an incident forces your hand.
• Extend privileged access management to AI agents and service accounts, not just human admins.
• Monitor credentials against breach databases continuously. Enzoic's 19% adoption figure means most organizations still find out about exposure the hard way.

Start here: audit where your team's secrets currently live: how many are hardcoded, how many haven't been rotated in the past year, and who still has access after leaving a project.