Build1 distinct publisher3 min readPublished
Seven entries went into CISA's known-exploited catalog on 2 September. The 5 September fix date covers four of the six products named, and the two AI-stack entries are the ones most teams cannot locate in a dependency graph.
The Engineer · Build desk

Compiled by The EngineerSomething wrong?How this is made
Authentication that turns on how a path is read gets decided twice: once by whatever terminates TLS in front of the service, and once by the application's own router. CISA describes the three software entries as a single attack path, starting with a crafted request to a public LiteLLM instance, to a service using Starlette, or to Kestra [5], and defeating bearer token validation, URL path interpretation, or suffix checks for `/configs` [6]. What follows the bypass is Model Context Protocol tool use, protected operations, or arbitrary workflow creation [7]. In Kestra, it is script execution plugins running code with root privileges [8].
Seven entries map onto six products, so at least one product contributes two of them [24], and the dev.to summary of the catalog entry ties those flaw classes to the three products as a group rather than one by one [6]. If you need to know which CVE describes the version you run, the advisories listed alongside the entry are where the mapping lives: GHSA-7488-6r32-c95q for LiteLLM, GHSA-86qp-5c8j-p5mr for Starlette, GHSA-5vc5-wxxq-3fjx for Kestra [18].
The window is shorter than the calendar suggests. From the 2 September entry [1] to the 5 September deadline [3] is three days [21], and 5 September 2026 is a Saturday, which leaves Thursday and Friday as the only full working days inside it [22].
CISA's mitigation list asks for the patched version plus external exposure moved behind a VPN, an allowlist or an authentication proxy, and then for restrictions on shell and script launches, token creation and outbound communication from management planes [16]. For a workflow engine, that second half constrains the feature you deployed the thing for. The version of it that survives contact with production is a non-root service account and an egress policy, not a ban on script plugins.
A model gateway holds provider keys and holds tool wiring, and the listed impacts pair those two: theft or tampering with API keys, administrator tokens, configurations and artifacts, then lateral movement into internal networks, cloud environments and delivery pipelines [19].
Whether Saturday is an incident date or a maintenance ticket depends on the three conditions the write-up states: the vulnerable version is reachable, the patch is not applied, and nothing else blocks management operations or code execution after a bypass [15]. Most of the seven need no user interaction [14], so there is no user report to wait for. On Switchvox, the SQL produced by concatenating `PhoneIP` into a PostgreSQL statement and abusing `PROGRAM` lands in `db-quirks.log` [9][17], a filename that is about to teach someone a lesson about naming. Elsewhere the signal is shells or unexpected child processes under Java, Python, PostgreSQL or VPN appliance parents [17].
One of these entries is a product you chose, and one is a framework something of yours imports: CISA's phrasing for the second target is "a service using Starlette" [5]. Four of the six named products carry the dated deadline, and the two AI-stack libraries do not [23], which makes that half of the work self-scheduled. If producing the list of services that import Starlette, and the routes your gateway answers on a public address, takes longer than an afternoon, then three days was never the binding number.
Ranked by verification strength, evidence, and original report placement.
CISA added seven vulnerabilities to its Known Exploited Vulnerabilities (KEV) catalog on 2026-09-02, affecting LiteLLM, Starlette, Kestra, JFrog Artifactory, Sangoma Switchvox and SonicWall SMA 1000.
CISA set a remediation deadline of September 5, 2026 for Kestra, Artifactory, Switchvox and SonicWall.
The CVEs associated with the entries are CVE-2026-59822, CVE-2026-48710, CVE-2026-49869, CVE-2026-82329, CVE-2026-9586, CVE-2026-83548 and CVE-2026-83549.
The described attack path begins with an attacker sending a crafted request to a public LiteLLM instance, a service using Starlette, or Kestra.
Flaws in bearer token validation, URL path interpretation, or suffix checks for /configs allow the attacker to bypass authentication boundaries.
After the bypass the attacker utilises Model Context Protocol (MCP) tools, reaches protected operations, or creates arbitrary workflows.
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · September 2, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
security
Microsoft's own incident data says the AI gateway is now the credential store1 distinct publisher
security
A poisoned Nx Console build rode VS Code's auto-update into GitHub's own repositories1 distinct publisher
security
CISA gives federal agencies three days to patch a 2023 ownCloud auth bypass4 distinct publishers
security
Attackers hid a cryptominer inside a LiteLLM MCP config test that reported success1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Authoritative upstream, single relay
The specifics that matter — seven entries, the 5 September date, the PhoneIP value dropped unescaped into a PostgreSQL statement, db-quirks.log as the place the query lands — all reach us through one dev.to write-up of CISA's notice. To its credit it shows its work, naming the three GitHub advisories, SonicWall's bulletin and Horizon3.ai's Switchvox analysis, so the chain back to primary material exists. Nobody in our coverage has walked it, and the one place the sourcing goes vague, the pre-authentication SonicWall chain, is exactly the place the severity claim rests.
Exploitation asserted, exposure uncounted
Presence in the exploited-vulnerability catalog is real-world usage evidence of a sort — the bar for entry is exploitation, not disclosure — and the Switchvox description of shells and exfiltration reads like hands-on activity rather than a proof of concept. What is entirely absent is scale: no number of internet-facing LiteLLM gateways, no Artifactory patch telemetry, no estimate of how many of the four deadline-bound products were still unpatched on the Saturday. The direction is documented; the magnitude is not.
Chains written fuller than confirmed
Modest overstatement, and it comes from form rather than adjectives. Each attack path is laid out as a clean sequence — crafted request, bypass, MCP tools, root — which reads as observed practice, when only Switchvox actually has post-exploitation behaviour attached to it. The framing that an LLM gateway now sits in the exploited catalog is not hype; it is the genuinely new fact, and if anything the piece underplays how awkward Starlette is to locate in a dependency tree.
Agency notice, vendor-shaped detail
The originator sells nothing: a government catalog entry with a compliance date is about as low-motive as security material gets, and the dev.to author gains only readers. The pressure sits one layer down, in the details that make the story vivid — the Switchvox teardown comes from a firm whose business is offensive testing, and the appliance narrative comes partly from SonicWall's own advisory, which decides how much of its own severity to foreground. Neither is a reason to disbelieve any of it; both are reasons the emphasis skews toward appliances over libraries.
Solid on record, silent on scale
Confidence splits by claim type. The record-level facts — what was added, when, which products face the Saturday date — are the kind of thing a single relay gets right, and the calendar arithmetic behind the two-working-day window is ours and checkable. Everything about breadth and in-the-wild behaviour rests on one unattributed sentence and no independent reporting, which is why this sits mid-range rather than high.