☀️ Good morning. Here's everything that happened in cybersecurity yesterday — in under 5 minutes.
The Justice Department pulled $400 million out of TikTok and ByteDance to close its child-privacy case, with the last $100 million riding on a court vacating an old Musical.ly consent decree; Corey Quinn argued in The Register that AWS's quarantine policy for leaked keys still leaves an attacker room to assume other roles, read secrets and lock storage; an independent scorecard graded five frontier AI labs on six control practices and found almost no public plan for containing a rogue model; Nozomi Networks showed at Black Hat that an unprotected Time-Sensitive Networking protocol can be used to nudge physical processes in OT plants; and new Black Kite data put 73% of disclosed ransomware and extortion incidents on mid-market companies. The through-line: the controls everyone assumes are load-bearing keep turning out to be partial.
🔥 Top Stories
01 — TikTok and ByteDance to Pay $400 Million to Settle DOJ Child-Privacy Lawsuit
Policy & Government
The Department of Justice announced on August 21, 2026 that TikTok and parent company ByteDance will pay $400 million to resolve a 2024 lawsuit alleging violations of federal children's privacy law, one of the largest COPPA-related recoveries the U.S. government has ever secured. The money moves in two tranches: $300 million on entry of the settlement, and $100 million later. That second payment is not tied to a calendar date. It is released only when a court vacates the earlier Musical.ly consent decree that has governed the company since the acquisition.
The contingent tranche is the part worth studying. Regulators structured a nine-figure payment around clearing an old order, which tells you they treat legacy consent decrees as live leverage rather than closed history. If your organization operates under a prior regulatory order, or inherited one through an acquisition, assume it is still being read. Confirm who owns compliance evidence for it and whether the obligations it created are actually running today.
02 — Corey Quinn Says AWS's Leaked-Key Quarantine Leaves Attackers Too Much Room
Cloud Security
When AWS Security finds an exposed credential in public sources, it programmatically attaches the AWSCompromisedKeyQuarantine policy to the offending IAM principal. The key stays active. The deny-list mostly targets fraud and cost abuse: buying savings plans, reading and deleting S3 object data, modifying Lambda functions, reading the CloudTrail audit log. Writing in The Register, Corey Quinn walked through what the policy leaves open and argued the deny-list blocks the obvious bill-running abuse while leaving an attacker enough room to gut the account from the inside, including by assuming other roles that the quarantine never touches.
AWS's counterargument is in its own documentation: deactivating a credential outright breaks every workload depending on it until the customer rotates. Quinn concedes that trade-off exists and disputes where the line sits. AWS has not publicly responded. The defender takeaway does not depend on who wins: treat a leaked key as a full compromise, deactivate or delete it, then work the blast radius through CloudTrail for assumed roles and new resources, check S3 Object Lock and retention changes, verify backups and snapshots, and confirm logging was never stopped. Move CI and third-party integrations to short-lived STS credentials and workload identity federation so the next leaked key expires in an hour.
03 — Frontier AI Labs Have Few Public Plans to Contain a Rogue Model, Study Finds
Artificial Intelligence
Guidelight AI Standards, led by chief scientist Steven Adler, a former OpenAI safety researcher, graded Anthropic, Google, Meta, OpenAI and xAI against six foundational control practices. On Guidelight's 0-to-5 scale, no lab exceeded a 3 on any single practice, and most scores landed at 2 or below. Overall: Anthropic and OpenAI tied at C+ (2.50), Google D+ (1.50), xAI D- (0.83), Meta F (0.67). On the specific practice of having a documented plan to contain a rogue model, OpenAI scored a 3, Google a 2, xAI a 1, and Meta and Anthropic both scored 0.
Read the caveat before you read the grades: this measures what each lab has disclosed publicly, not what it does internally. A lab could have a serious containment program and score badly for saying nothing about it. That distinction matters more for buyers than for the labs. If you are putting a frontier model inside a workflow that touches production, public disclosure is the only control evidence you can actually verify, and right now the evidence is thinnest exactly where the consequences are worst.
04 — Unprotected TSN Protocols Could Disrupt Physical OT Processes, Nozomi Research Shows
Industrial Control Systems
At Black Hat USA, Nozomi Networks presented research on Time-Sensitive Networking, the family of Ethernet amendments that guarantees an emergency-stop signal arrives within microseconds even on a congested shared network. The team studied one protocol, CC-Link IE TSN, and showed that when it runs without its full set of cyber controls, an attacker already inside the network can tamper with the process variables the protocol is supposed to carry safely. The most important entry point, researcher Cremona said, is the management interface of the TSN switch. The team also reported previously undiscovered bugs in Phoenix Contact TSN switches, and a firmware patch is already available.
This is lab work and a conference demo, not an active campaign, and no victims are named. Two disclosures belong in the read: Mitsubishi Electric, which owns CC-Link IE TSN, completed its roughly $1 billion acquisition of Nozomi in January 2026, six months before the session, and Nozomi says it operates independently. The action is unglamorous and immediate. Patch OT network device firmware, starting with the Phoenix Contact fix, segment TSN traffic away from anything that lets an intruder reach a switch management interface, and ask your TSN vendors what integrity protections ship today versus what is on the roadmap.
05 — Mid-Market Companies Absorbed 73% of Disclosed Ransomware Incidents Since 2023
Ransomware
Black Kite analyzed 13,336 publicly disclosed ransomware and data-extortion incidents with known victim revenue across North America and Europe between January 2023 and June 2026. Companies with annual revenue between $10 million and $1 billion accounted for 73% of them, and that share never moved outside a 72% to 75% band across the entire period. This is not a spike or a shift in targeting. It is a stable, three-and-a-half-year baseline.
The stability is the finding. Mid-market firms sit in the gap where the data is worth extorting but the security program is usually one team, sometimes one person, and rarely a 24/7 detection capability. If you are in that revenue band, the planning assumption should be that you are the median target rather than an outlier, which changes what you fund first: tested offline backups, enforced MFA on remote access and identity providers, and a rehearsed incident runbook that does not assume a vendor will answer at 2 a.m.
📊 By The Numbers
$400 million — Total TikTok and ByteDance will pay the DOJ, split into $300 million on entry and $100 million released only when a court vacates the Musical.ly consent decree.
0 — Guidelight's rogue-model containment-plan score for both Meta and Anthropic; OpenAI scored 3, Google 2, xAI 1, on a 0-to-5 scale.
2.50 — The top overall grade in the frontier-lab scorecard, a C+ shared by Anthropic and OpenAI; Meta finished at 0.67, an F.
73% — Share of 13,336 disclosed ransomware and extortion incidents that hit companies with $10 million to $1 billion in revenue, January 2023 through June 2026.
$1 billion — Roughly what Mitsubishi Electric paid to acquire Nozomi Networks in January 2026, six months before Nozomi presented research on Mitsubishi's own CC-Link IE TSN protocol.
⚡ The Signal
Four of yesterday's five stories describe the same failure mode from different angles: a control that exists, that everyone treats as load-bearing, and that turns out to cover less than its name implies. AWS quarantines a leaked key, and the key still works. TSN guarantees deterministic delivery, and determinism was never integrity. A frontier lab publishes a safety framework, and the containment practice inside it scores a zero. In each case nobody lied. The control did exactly what it was scoped to do, and the scope was narrower than the assumption built on top of it.
That gap is where defenders get hurt, because partial controls are worse than absent ones. An absent control shows up in a gap analysis. A partial control shows up as a green check. Quinn's argument is not really about AWS policy syntax; it is that "quarantined" reads like "contained" to the person triaging the alert at 3 a.m., and the difference between those two words is an entire incident. The same is true of a TSN switch whose management interface is reachable, and of a model card that documents evaluations but not what happens when one fails.
The practical move is to stop grading controls by whether they exist and start grading them by what they explicitly do not cover. Read the deny-list, not the policy name. Ask which protocol layer the integrity protection actually sits on. Ask a vendor which of its six practices scored zero and why. The TikTok settlement makes the same point from the regulator's side: an old consent decree that everyone had mentally filed as resolved turned out to be worth $100 million, because someone in government was still reading it.
🔍 What You May Have Missed
The CyberSignal published four articles Sunday; these three ran over the weekend and are still worth your time.
North Korean Hackers Tied to the Rust crates.io Supply-Chain Attack, Wiz Says — Wiz linked the compromise of three Rust crates to recent North Korean supply-chain operations via command-and-control infrastructure shared with the Mastra and axios npm attacks. Strong lead built on overlap, not a confirmed verdict.
iAuthFlow V2 Phishing Kit Plants Attacker Passkeys to Survive Password Resets — Abnormal documented a $10,000 toolkit that registers an attacker-controlled passkey on a compromised account. Because a passkey is a credential rather than a session token, it survives the password reset and session revocation that normally evict a phisher.
Microsoft Defender's Signed BTR.sys Driver Can Be Weaponized in the Kernel, Check Point Finds — Defender's own legitimately signed remediation driver can be turned into a kernel primitive for arbitrary file and registry changes across Windows 7 through Windows 11 25H2, with no external driver imported.
📅 What to Watch
GitLab CVE-2026-19478 Is Under Active Exploitation — The Hacker News independently confirmed watchTowr's report that this CVSS 9.4 unauthenticated GraphQL code-injection flaw is being exploited within days of disclosure. Unpatched self-managed instances are the exposure; patch this week.
Microsoft Entra ID CVE-2026-69836 Was Exploited Before the Fix — A maximum-severity RCE in Entra ID was exploited in the wild before Microsoft fixed it on its own infrastructure. There is no patch to install, which is not the same as nothing to check; expect follow-on detail on scope.
AWS Adds Rule Hit Counts to Network Firewall — Stateful rules in custom and managed rule groups now report whether they are matching traffic, enabled by default at no extra Network Firewall cost. A cheap way to find the rules you think are protecting you and are not.
Week in Review: Azure Tenant Records and Medusa's 500+ Victims — Records allegedly stolen from Azure tenants and a Medusa ransomware count past 500 organizations are both still developing, alongside research on bypassing Windows 11 defenses without physical access.
Stay sharp. Stay ahead.
Till next time,
The CyberSignal Team