Forescout's AI harness: $2.58 per candidate, $9.44 per confirmed bug (derived), and 24 of 54 reported fixed
Forescout says its AI harness found 99 TCP/IP stack vulnerabilities; the $2.58 per finding divides $935 by 362 candidates, which is $9.44 per confirmed bug (derived). Its own table marks 24 of the 54 reported as fixed (derived sum); Siemens' advisory gives an update for 6 of 23 listed rows.
By Parminder Kumar Sharma · LinkedIn · · 26 min read
$2.58 per finding divides by 362 candidates, not by 99 bugs
Forescout's research team says its AI harness found 99 vulnerabilities in about 90 hours in ten TCP/IP stacks, the networking code inside a device, and prints a cost of $935 for the static analysis stages, an average of $2.58 per finding. The report also prints the divisor: 362 candidate findings. The same $935 over the 312 unique candidates is $3.00, and over the 99 confirmed issues it is $9.44 (derived: 935 / 362 = 2.58, 935 / 312 = 3.00, 935 / 99 = 9.44). Add the report's own estimate for the dynamic stage, $290 to $1,380, and the model spend is $1,225 to $2,315 (derived: 935 + 290 and 935 + 1,380), which is $12.37 to $23.38 per confirmed issue (derived: divided by 99).
What that does not establish. It does not show the $2.58 is wrong: the report states what it divides. It does not price the work of creating the harness, which the report excludes, and it does not price the manual research the report compares against, because no figure is printed for that side of the 100 times claim. The 99 also includes 33 vulnerabilities that Forescout had found before; the new ones number 66.
The second number is the status. The report says it reported 54 of the 66 new findings, and its own table marks 24 of them Fixed (derived: 19 open source plus 5 for Nucleus NET, from the per stack counts). Its summary sentence says "only five have been fixed to date", and that is the count for the 34 proprietary findings, not for all 54. The Siemens advisory of 13 October 2026, which the report links as the fix for Nucleus NET, gives an update for 6 of the 23 product rows it lists and says no fix is planned for the other 17 (derived: counted from the advisory).
Whose figures these are. Forescout sells security products and promotes its own Assume Autonomy framework in the same report. This is the vendor's own research, it benefits from a story in which AI makes finding bugs cheap, and every count below is Forescout's count, apart from the Siemens advisory and the GitHub records, which are independent of it. The report describes a study of ten stacks: seven from its 2020 to 2021 Project Memoria and a similar study by other researchers, and three new ones, lwIP, wolfIP and one binary it calls S3.
What Forescout says it created, and the ten stacks it ran it on
Forescout says it tried open harnesses first, got few meaningful results, and wrote its own, mixing fixed rules with AI steps. Stage 0 prepares binary targets, using Ghidra to produce a common representation. Stage 1 runs static analysis: Semgrep on source code and Joern, with 32 rules, for data and control flow. Stage 2 has AI agents judge whether the flagged code can be reached with attacker input, and stage 3 has agents judge whether what remains is exploitable. Stage 4, dynamic verification, has an AI agent compile or emulate the target and try to produce a proof of concept script that triggers the flaw; if it triggers, the issue is confirmed, otherwise it is discarded. Stages 0 to 3 are automated; stage 4 is set up so a researcher can step in, and the report says the approach is not fully autonomous and that the 90 hours include human validation.
On the models, Forescout's account is this: it used Anthropic's Claude family because of its earlier evaluations and what it calls not too restrictive guardrails, with Sonnet 4.6 for the first three stages and Opus versions between 4.6 and 5 for dynamic verification. It says Claude helped write the 32 Joern rules from Forescout's earlier reports, exploits and rules, and that the rules were then manually reviewed. It also says it rarely met guardrail refusals with older models and met them more on recent ones. That is the vendor's description of its own method; no independent record of it was read.
The ten stacks as the report's tables print them. Counts of Memoria+ vulnerabilities are from its first table; candidates, confirmed issues and precision from its second. Total row: the 82 is a sum of the ten rows (derived) and the pooled rate is 99 / 312 (derived).
| Stack and artifact type | Memoria+ vulnerabilities | Candidates, then confirmed | Precision |
|---|---|---|---|
| S2 (binary) | 14 | 23, then 16 | 70% |
| FNET (source) | 6 | 9, then 6 | 67% |
| wolfIP (source) | 0 | 5, then 3 | 60% |
| Ethernut (source) | 5 | 20, then 11 | 55% |
| Contiki (source) | 14 | 40, then 17 | 43% |
| S3 (binary) | 0 | 32, then 12 | 38% |
| lwIP (source) | 0 | 13, then 4 | 31% |
| PicoTCP (source) | 11 | 44, then 12 | 27% |
| Nucleus NET (binary) | 20 | 53, then 10 | 19% |
| S1 (binary) | 12 | 73, then 8 | 11% |
| Total | 82 (derived sum) | 312, then 99 | 42% printed; 31.7% pooled (derived) |
The four binaries are S1, S2, S3 and Nucleus NET; the other six are source code. The report names the vendor of none of S1, S2 and S3, because disclosure is still going on. It adds that one of the three had its end of general support in December 2025 and will reach the end of extended support in September 2027, without saying which.
From 362 candidates to 24 fixed
What each count is, and where it stops being checkable. Sources: the report's text and tables; derived values are this site's arithmetic.
| Count | What it is, per the report | What it does not tell you |
|---|---|---|
| 362 | Candidate findings from the static stages across ten stacks, for roughly $935 | Which are real; the report prints a range of $0.95 to $4.16 per candidate by stack, and the table behind it was not saved as text |
| 312 | Unique candidates after deduplication (50 fewer, derived) | Why 50 were duplicates, or whether duplicates were charged for |
| 99 | Confirmed issues at the dynamic stage: 33 rediscovered of the 82 in Memoria+ (40%; 33 / 82 = 40.2%, derived) and 66 new | The 82 does not match the 71 printed elsewhere in the report; see the differences below |
| 66 | New: either missed in the original analysis or in stacks not analysed before | A count of distinct root causes, or how many were byproducts; the report says only that the logic bugs were byproducts |
| 54 | Reported to vendors and maintainers, because the report says it believed they still affected the latest versions (20 open source, 34 proprietary) | Why 12 were not reported (derived: 66 - 54); the report does not list them |
| 24 | Findings whose stack the report's status table marks Fixed (derived: 19 + 5) | That a device is patched. Fixed here is a status for a stack, and the 29 still being disclosed are not counted |
The cost and time claims, one at a time
The report's headline claims are about time and money, so each is set against what the page prints. The table keeps the claim, the printed support and the gap.
Cost and time claims, as the report prints them and what each leaves out. Sources: the report's text and its comparison table; derived values are this site's arithmetic.
| Claim | What the report prints | What it leaves out |
|---|---|---|
| $935, $2.58 per finding | Static analysis stages, 362 candidates; stages 1 to 3 took about ten hours of wall clock time for all ten targets (derived check: six source targets under 40 minutes and four binaries at about 90 minutes is 600 minutes) | Cost per confirmed issue: $9.44 over 99, or $14.17 over the 66 new (derived: 935 / 66) |
| $290 to $1,380 for dynamic analysis | An estimate: the report has no precise time or cost measurements for stages 0 and 4. It estimates $15 to $30 per open source target and $50 to $300 per binary | A measured figure. The range is the sum of those per target estimates (derived: 6 x 15 + 4 x 50 = 290 and 6 x 30 + 4 x 300 = 1,380) |
| $1,200 to $2,300 of model spend | Rounded from 935 plus the dynamic range, which is $1,225 to $2,315 (derived) | Anything outside model spend: researcher time for validation, tools, emulation hardware |
| 100 times cheaper | The report says the model spend is likely 100 times less than a senior researcher doing the manual code analysis; the key findings word it as against the previous analysis | Any figure for the manual side. Taken at face value the claim implies $120,000 to $230,000 of manual cost (derived: 100 x 1,200 and 100 x 2,300), a number the report does not print |
| 9 months to 90 hours | Memoria's code analysis took about 9 months; this study took about 10 hours of AI analysis plus about 80 hours of dynamic validation, human validation included | A like for like unit. Memoria took 18 months over 14 stacks in five phases; this study ran 10 stacks in one phase. For scale only, 9 months at 40 hours a week is about 1,560 hours (derived, this site's assumption), 17.3 times 90 |
| Marginal cost | The report says the comparison excludes designing the harness, writing and refining the rules, creating the emulation set up and refining prompts | How long or how much that took |
The report's own conclusion is that once an actor has such a harness it can find this class of bug quickly and cheaply, and that this is true for attackers and defenders alike. It also says the three things most such work relies on, a cooperative frontier model, a sizeable budget and source code, are not always available to attackers, and that binaries cost about six times more and take about twice as long. Both statements are in the report. Which of them applies to a given attacker is not something it measures.
Precision: the 42% is an average of ten averages
The precision table prints 42% as the average. The ten per stack rates, recomputed from the candidate and confirmed counts, average 41.9% (derived), so 42% is their simple mean. Pooled over all candidates, 99 confirmed out of 312, the rate is 31.7% (derived). The gap exists because the stacks with the most candidates, S1 with 73, Nucleus NET with 53 and PicoTCP with 44, are the ones with the lowest precision.
The report says source code averaged 47% and binaries 34%. Pooled, source stacks come to 53 confirmed of 131 candidates, 40.5%, and binaries to 46 of 181, 25.4% (derived). The report says binary precision falls to 22% without S2, an outlier because its demo binary kept symbols and ELF headers; the mean of the other three rates, 11.0%, 37.5% and 18.9% unrounded, is 22.4% (derived), which matches the printed 22. In the report's words for the range, in the best case roughly one finding in three was wrong and in the worst nine in ten. Nucleus NET, the stack tied to the Siemens advisory, is a binary at 19%.
The status of the 54: what Fixed counts
The report's status table, as printed, with this site's sums (derived). The fix links are the ones the report's page carries.
| Stack | New findings and status | Fix links on the report's page |
|---|---|---|
| Contiki | 4, Fixed | 4 pull requests in a repository named contiki-ng |
| Ethernut | 1, No response | None |
| FNET | 3, Fixed | 1 pull request |
| PicoTCP | 5, Fixed | 1 pull request |
| wolfIP | 3, Fixed | 2 commits |
| lwIP | 4, Fixed | 3 commits |
| Open source, derived | 20 reported, 19 Fixed, 1 No response | 11 GitHub records |
| Nucleus NET | 5, Fixed | The Siemens advisory SSA-333035 |
| S1, S2, S3 | 7 + 10 + 12 = 29, Investigating | None |
| Proprietary, derived | 34 reported, 5 Fixed, 29 Investigating | 1 advisory |
| All, derived | 54 reported, 24 Fixed (44.4%) | 12 links for 24 Fixed findings |
The summary sentence. The report's executive summary reads 20 affecting open source stacks and 34 affecting proprietary stacks, with the five fixed in a bracket after the 34. Read as printed, that is the proprietary count, and the table agrees: Nucleus NET, 5. Read quickly, it can look like five of 54. Help Net's relay does not repeat the five; it says several projects have issued fixes and that investigations remain open for three stacks whose names are withheld. Open source is therefore 19 of 20 fixed (95%, derived) and proprietary 5 of 34 (14.7%, derived), and the 24 of 54 sits between them.
Fixed is a label on a stack. For six stacks the status is Fixed and a link follows; the report does not say whether a fix has reached any product or device that embeds the stack. The Nucleus NET row is the clearest case, and the next section sets it against Siemens' advisory. The report says in its closing section that patching embedded devices is notoriously difficult and that fixes must still travel through the supply chain.
The appendix does not match the status table. The report's appendix says it lists the vulnerabilities that have already been fixed. It prints 25 rows: 19 for the five open source stacks marked Fixed, 5 for Nucleus NET, and 1 for Ethernut, whose status is No response (derived: 4 + 3 + 4 + 5 + 3 + 1 + 5 = 25). The Ethernut row carries CVE-2026-120596, a denial of service in the TCP options parser of Nut/OS. The same appendix says CVE IDs for the open source findings were requested from MITRE and are awaited, so one is printed and nineteen are not: the other rows carry Forescout's own FSCT identifiers. The report does not explain why a stack with no response has a CVE ID, and this site did not read the CVE record. PicoTCP's five rows run FSCT-2026-0017 to 0020 and 0022; 0021 is not printed.
S1 is also a status without a party. The report says the S1 stack it analysed sat in the firmware of a discontinued wireless access point made by a company that no longer exists, and that the harness found seven vulnerabilities specific to those devices, including a buffer overflow in firmware upload and administrator credentials that cannot be changed. S1's status is Investigating. The report does not say who is investigating, and says that no vendor will be able to provide a patch for unsupported products like this one.
What the Siemens advisory prints against the report's Fixed
Siemens ProductCERT advisory SSA-333035, version V1.0, was published on 13 October 2026 and acknowledges two Forescout researchers for coordinated disclosure. It lists five CVEs and 23 product rows. It does not name Nucleus NET, any TCP/IP stack, or AI: the link between these five CVEs and Nucleus NET is the report's, which says Siemens was the first to identify vulnerable devices and that controllers in the APOGEE, Desigo and TALON lines run Nucleus NET. The report's appendix repeats the advisory's five descriptions word for word.
The advisory's product rows grouped by family, as printed, with this site's counts (derived). Source: Siemens advisory SSA-333035, as read on 16 November 2026.
| Rows | CVEs and affected versions | Remediation printed |
|---|---|---|
| APOGEE MBC and MEC, BACnet and P2 Ethernet (4 rows) | CVE-2026-102073 to 102076; all versions | No fix is planned. The advisory calls these legacy devices, replaced by the APOGEE PXC Modular and Compact series |
| APOGEE PXC Compact and Modular, BACnet and P2 Ethernet (4 rows) | CVE-2026-102073 to 102076; below V3.5.7 (BACnet) or below V2.8.21 (P2 Ethernet) | Update to V3.5.7 or later, or V2.8.21 or later |
| TALON TC Compact and Modular, BACnet (2 rows) | CVE-2026-102073 to 102076; below V3.5.7 | Update to V3.5.7 or later |
| Desigo PXC and PXM controllers (13 rows) | CVE-2026-89208 only; all versions from V2.3 | No fix is planned; apply network segmentation and put the controllers behind properly configured gateways or firewalls |
| All rows, derived | 23 rows (4 + 4 + 2 + 13) | 6 rows with an update, 17 with no fix planned (4 + 13) |
What the advisory says about exposure. For the four service CVEs, 102073 to 102076, the mitigation is to disable the FTP and web service, and the advisory notes that the service is disabled by default on APOGEE, Desigo and TALON products. Exposure for those four therefore depends on a service the advisory says ships off, and the advisory does not say how many installed controllers have it on. The note names Desigo too, though no Desigo row lists those four CVEs. The fifth, CVE-2026-89208, an out of bounds read, has no service to switch off; Siemens recommends segmentation, and 13 Desigo rows have no fix planned.
Scores, as printed. The advisory's headline scores are CVSS v3.1 8.2 and v4.0 8.8. By CVSS v3.1 base score the five are 5.3, 8.2, 7.5, 7.5 and 7.5, which by the v3.1 rating bands makes four high, one medium and none critical (derived). This site recomputed each v3.1 score from its printed vector and all five match. Forescout's 52% critical or high is a different set, and the severity scores for its other findings are not printed. The advisory's summary speaks of denial of service and availability, while the entries for CVE-2026-89208 and CVE-2026-102073 describe reading memory; the advisory does not reconcile the two.
So the difference is this. The report's table says Nucleus NET, 5, Fixed. The independent record says that, for the products it lists, 6 rows have an update and 17 do not, that the update rows belong to APOGEE PXC and TALON lines, and that no row says Nucleus NET. Both can be true: a stack vendor may have a fix while most product lines that embed it have none planned. The label Fixed, on its own, does not tell an owner of a Desigo controller which of those applies.
The GitHub records the report links as fixes
The report links eleven GitHub records as fixes or advisories for the open source stacks. This site read each as GitHub prints it. The table gives each title or commit message and its dates; the report says Forescout reported the findings and says which links fix them, and this site does not claim any record fixes any finding. Only two of the eleven name Forescout, and none mentions a CVE.
The eleven GitHub records, as GitHub prints them (UTC dates). Pull requests show the date opened and the date merged; commits show the author date and the committer date. One pair of quotation marks in the tftp message is omitted. Source: the saved GitHub records.
| Record | Title or commit message, as printed | Dates, UTC |
|---|---|---|
| Contiki pull request 3256 | ip64: check the IPv6 payload length against the payload | Opened 4 Sep 2026; merged 7 Sep 2026 |
| Contiki pull request 3257 | ip64: bound the DHCPv4 option walk | Opened 4 Sep 2026; merged 16 Sep 2026 |
| Contiki pull request 3258 | roll-tm: bound the sequence lists in an ICMPv6 message | Opened 4 Sep 2026; merged 27 Sep 2026 |
| Contiki pull request 3282 | rpl: fix DIO output buffer overflow from stale uip_ext_len | Opened 8 Oct 2026; merged 13 Oct 2026 |
| FNET pull request 20 | Fix three remotely-triggerable memory-safety/DoS issues (#19, FSCT-2026-0023/0024/0025) | Opened 7 Sep 2026; merged 7 Sep 2026 |
| PicoTCP pull request 81 | Fix five memory-safety issues from Forescout (Vedere Labs) disclosure (FSCT-2026-0017/0018/0019/0020/0022) | Opened 8 Sep 2026; merged 8 Sep 2026 |
| wolfIP commit 8bfdb80 | Fix Pico USB network buffer dimensions and bounds | Authored 15 Jun 2026; committed 30 Jun 2026 |
| wolfIP commit b9e84e3 | F-5784: strip IP options before ESP unwrap so IHL>5 ESP packets are handled | Authored and committed 4 Jun 2026 |
| lwIP commit 81bfe39 | mqtt: fix infinte loop via crafted publish for some pbuf configurations | Authored and committed 26 Oct 2026 |
| lwIP commit 89b823c | smtp: fix heap buffer overflow when application provides oversized to/from/subject | Authored 2 Jun 2026; committed 27 Oct 2026 |
| lwIP commit 9670da2 | tftp server: fix sys_timeout pool exhaustion via rejected RRQ/WRQ requests | Authored 2 Jun 2026; committed 27 Oct 2026 |
Who is named. FNET pull request 20 and PicoTCP pull request 81 name Forescout and its FSCT identifiers. Three Contiki pull requests, 3256, 3257 and 3258, thank two reporters by name; the same two names are the ones Siemens credits as Forescout staff, and this site does not print them. Pull request 3257 also credits a verification tool group that reported the same parser defect independently. Pull request 3282 credits no one. The lwIP and wolfIP commits name no reporter.
Dates. Seven of the eleven records are dated before the Siemens advisory of 13 October: the FNET pull request merged 36 days earlier and the PicoTCP one 35 days earlier (derived), Contiki 3256, 3257 and 3258 between 16 and 36 days earlier (derived), and the two wolfIP commits, committed on 4 and 30 June, the earlier one 131 days before the advisory and 161 days before the report's page date (derived). Contiki 3282 was merged on 13 October itself. The lwIP commits are dated 26 and 27 October, 13 and 14 days after the advisory (derived), and two of them were authored on 2 June and committed on 27 October; the records print both dates and say nothing about the gap. The report prints no date for when it reported any finding, so these dates neither confirm nor contradict the report's account.
Records and findings are not paired. The report links per stack, not per finding. lwIP has four findings and three links; wolfIP has three findings and two. The three lwIP commit subjects name MQTT, SMTP and TFTP, the protocols of three of the four lwIP rows in the appendix, and the fourth row, in TCP receive code, has no commit that names it. For Contiki, the report links four pull requests and prints four rows. The description of pull request 3282 is a buffer overflow when a node replies to a DIS message, while the nearest appendix row, FSCT-2026-0015, describes parsing a received DIO message; the report does not say which pull request fixes which row, and this site does not say whether they match. The report also names Contiki v3.0 for its analysis and links pull requests in a repository named contiki-ng without explaining the relationship, and it names PicoTCP v1.7.0 in one table and PicoTCP-NG in its appendix.
Two descriptions that sit beside the report's. Pull request 3256 describes a length check in the IPv6 to IPv4 translation code and says that no caller reaches it today, because the caller already drops such packets. The report's FSCT-2026-0016 describes the same translator and says any node on the gateway's internal side can trigger it unauthenticated. The report does not map the two, and neither does the pull request. For PicoTCP, pull request 81 says the fix for FSCT-2026-0018 is a change to a shipped example plus a documented buffer size contract, with a library change suggested but not made, and describes FSCT-2026-0020 as low practical severity, reported for hardening. So a Fixed status for PicoTCP's five findings includes one fixed in an example and one the maintainers call hardening.
Where the records differ
The report, Help Net, the Siemens advisory and the GitHub records disagree, or leave gaps, in several places. The table lists them and resolves none.
Differences and gaps between the saved pages. This site resolves none of them. Sources: the pages named in each row.
| Point | One account | Another account |
|---|---|---|
| Vulnerabilities in the comparison | The first table: 71 across the stacks re-analysed here | The text and the stack table: 82 in Memoria+ (14 + 5 + 6 + 11 + 12 + 14 + 20, derived), of which 33 were rediscovered. The 11 difference is not explained |
| Which set is 52% critical or high | The key findings line says of the vulnerabilities, which reads as the 99 | The FAQ says of the 66 new, and Help Net says of those newly discovered. 34 of 66 would print as 52% (derived: 51.5%). The severity chart was not saved as text |
| The 29% in core components | Key findings: 29% of the vulnerabilities affect core components | The FAQ says seventeen of the new ones affect core IP, TCP, UDP, ICMP or Ethernet code, which is 25.8% of 66 (derived); 29% of 99 would be about 29 (derived: 28.7) |
| The report's date | The visible text prints none; its page data says 12 November 2026, 12:00 UTC | Help Net dated its relay 16 November 2026, 4 days later (derived) |
| How many are fixed | The report's summary sentence puts a count of five fixed in a bracket after the 34 proprietary findings | Its table: 24 of 54 Fixed (derived); Help Net: several projects have issued fixes. The appendix lists 25 rows as already fixed, one for a stack with no response |
| What Fixed means for Nucleus NET | The report's status: Fixed, linking to the advisory | The advisory: update rows for 6 of 23, no fix planned for 17, and no mention of Nucleus NET |
| Open source CVE IDs | The appendix: IDs requested from MITRE, awaited | The same appendix prints CVE-2026-120596 for Ethernut. No GitHub record mentions a CVE |
| Hundreds of products | Key findings: past disclosures suggest hundreds of products from dozens of vendors are likely affected | The body says Project Memoria identified more than 100 products from dozens of vendors. For this round only Siemens is named, and other vendors are investigating. No count of products was measured for this round |
| S1 | The report says the company behind the access point is defunct | The status table says Investigating, and does not say by whom |
The study's own limit. The report says the harness found bugs in memory safety and protocol parsing code and missed logic and design bugs, such as predictable initial sequence numbers, weak random numbers, DNS cache poisoning and TCP state machine problems, and that the logic bugs it did find were byproducts of other work. Of the 99, it counts 48 as memory corruption, 27 integer issues, 16 input validation and 8 control flow or logic (derived sum: 99). A harness tuned to one class of bug says little about the bugs it was not tuned for.
What the report and the records do not establish
What the saved pages state and do not state, as read on 16 November 2026. Sources: the report, the Siemens advisory, the GitHub records and the Help Net relay.
| Topic | Stated (by whom) | Not stated |
|---|---|---|
| Affected products | Siemens APOGEE, Desigo and TALON rows (advisory); other vendors are investigating (report) | Any other product; a count of products or devices |
| S1, S2, S3 | Binary stacks, 7, 10 and 12 findings, Investigating (report) | The stacks, their vendors, the dates reported, CVE IDs, whether anyone has fixed anything |
| Exploitation | Nothing: no saved page reports any in the wild | Whether any of the 66 has been used |
| Reach | Siemens: segmentation, and the FTP and web service off by default | Any scan, any count of reachable devices, any share with the service enabled |
| Reproduction by an attacker | The report says defenders and attackers alike could do it, and that models, budget and source code are not always available to attackers | Any test of an attacker with those limits; binaries cost about six times more (report) |
| The 100 times claim | Model spend $1,200 to $2,300 (report) | The manual cost it is compared with |
| Severity of the other findings | 52% critical or high of the 66 (FAQ); five Siemens scores | Scores or CVSS vectors for the rest; who assigned the 52% |
| Proof of concept | The dynamic stage generates proof of concept scripts (report) | Any script, payload or exploit: none is printed, and none is reproduced here |
| Fix to finding | Links per stack (report) | Which record fixes which finding, and the 12 unreported findings |
| Generalising | Forescout's choice of models and its guardrail experience (report) | Whether other models or other vendors' guardrails would give the same result |
| Replication | Nothing: the report is the only source of its counts | Any independent run of the harness |
What the advisory prints
The Siemens advisory is the only saved page that prints identifiers a defender can use. The report prints Forescout's own FSCT identifiers, CVE IDs and descriptions of each flaw, and it prints no proof of concept, although it says its dynamic stage generates them; none is reproduced here. The block below is copied from the advisory: for each of the five CVEs, the weakness class (CWE) and the CVSS v3.1 base score and vector. A defender uses it to match a scanner's finding to a CVE, to see that all five are scored as reachable over the network with no login and no user action (AV:N, PR:N, UI:N), and that CVE-2026-89208 is scored for confidentiality only (C:L), three for availability only (A:H) and CVE-2026-102073 for both (C:H, A:L).
CVE-2026-89208
CWE-125: Out-of-bounds Read
CVSS v3.1 Base Score 5.3
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:L/I:N/A:N
CVE-2026-102073
CWE-191: Integer Underflow (Wrap or Wraparound)
CVSS v3.1 Base Score 8.2
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:L
CVE-2026-102074
CWE-121: Stack-based Buffer Overflow
CVSS v3.1 Base Score 7.5
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
CVE-2026-102075
CWE-121: Stack-based Buffer Overflow
CVSS v3.1 Base Score 7.5
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
CVE-2026-102076
CWE-121: Stack-based Buffer Overflow
CVSS v3.1 Base Score 7.5
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
The advisory's four mitigations are, in this site's paraphrase: for CVE-2026-102073 to 102076, disable the FTP and web service, which the advisory notes is off by default on APOGEE, Desigo and TALON products; for CVE-2026-89208, apply network segmentation and put the controllers behind properly configured gateways or firewalls. The fixed versions are V3.5.7 for the BACnet rows and V2.8.21 for the P2 Ethernet rows of the APOGEE PXC series, and V3.5.7 for the two TALON TC rows.
What to do, in order
First Siemens' own steps and Forescout's, then this site's suggestions, each labelled.
Take this with you
In order
- Siemens, for APOGEE PXC Compact and Modular and TALON TC Compact and Modular controllers: update to V3.5.7 where the BACnet versions apply and to V2.8.21 where the P2 Ethernet versions apply (advisory SSA-333035).
- Siemens, for CVE-2026-102073 to 102076: disable the FTP and web service. The advisory says it is off by default; checking that it is off on each controller is this site's suggestion.
- Siemens, for CVE-2026-89208 on Desigo controllers, where no fix is planned: apply network segmentation and put the controllers behind properly configured gateways or firewalls.
- APOGEE MBC and MEC controllers have no fix planned, and the advisory calls them legacy devices replaced by the PXC series. Forescout advises isolating devices that cannot be patched, allowing only the traffic they need; this site suggests putting replacement on the plan.
- Forescout, for asset owners: prioritise by reachability, exploitability and operational impact; restrict services such as FTP, Telnet, IPv6 and web administration; shorten the time to patch where fixes exist.
- This site's suggestion: ask each supplier in writing whether any product embeds Nucleus NET, lwIP, Contiki, PicoTCP, FNET, wolfIP or Ethernut, which version, and which advisory covers it. The report recommends requiring SBOMs and support lifetimes in procurement.
- This site's suggestion: treat a fixed in the stack line as not yet a fixed device. Ask for the product release that carries the fix, and for the date.
- Forescout, for vendors and maintainers: expect more AI generated vulnerability reports, many of them low quality, and invest in triage and coordinated disclosure.
The question the counts leave open
The report counts 99 confirmed issues, 66 new, 54 reported and 24 marked Fixed; Siemens counts 23 product rows and 6 updates. The figure that decides risk, how many of those findings sit in a device that someone can reach, is in neither. The report prints no scan count, the advisory says its main service is off by default and gives no count of installations, and three of the ten stacks have no named vendor. How many of the 54 reach a device in the field, and who counts them?
Sources
- PrimaryThe report, From 9 Months to 90 Hours: Smashing TCP/IP Stacks: the counts, costs, status table, precision table and appendix. The vendor's own research. Visible text prints no date; page data gives 12 November 2026. Read by this site on 16 November 2026Forescout Researchaccessed 2026-11-16
- PrimaryAdvisory SSA-333035, V1.0, 13 October 2026: five CVEs, 23 product rows, mitigations, acknowledgement of Forescout. Independent of the report. Read by this site on 16 November 2026Siemens ProductCERTaccessed 2026-11-16
- PrimaryPull request 3256, ip64 IPv6 payload length check, merged 7 September 2026, as GitHub prints itGitHub (Contiki-NG)accessed 2026-11-16
- PrimaryPull request 3257, ip64 DHCPv4 option walk, merged 16 September 2026, as GitHub prints itGitHub (Contiki-NG)accessed 2026-11-16
- PrimaryPull request 3258, roll-tm sequence lists, merged 27 September 2026, as GitHub prints itGitHub (Contiki-NG)accessed 2026-11-16
- PrimaryPull request 3282, rpl DIO output buffer, merged 13 October 2026, as GitHub prints itGitHub (Contiki-NG)accessed 2026-11-16
- PrimaryPull request 20, three issues, FSCT-2026-0023 to 0025, merged 7 September 2026, as GitHub prints itGitHub (FNET)accessed 2026-11-16
- PrimaryPull request 81, five issues, FSCT-2026-0017 to 0022, merged 8 September 2026, as GitHub prints itGitHub (PicoTCP)accessed 2026-11-16
- PrimaryCommit 8bfdb80, Pico USB network buffer, authored 15 June and committed 30 June 2026, as GitHub prints itGitHub (wolfIP)accessed 2026-11-16
- PrimaryCommit b9e84e3, IP options before ESP unwrap, 4 June 2026, as GitHub prints itGitHub (wolfIP)accessed 2026-11-16
- PrimaryCommit 81bfe39, MQTT infinite loop, 26 October 2026, as GitHub prints itGitHub (lwIP)accessed 2026-11-16
- PrimaryCommit 89b823c, SMTP heap buffer overflow, authored 2 June and committed 27 October 2026, as GitHub prints itGitHub (lwIP)accessed 2026-11-16
- PrimaryCommit 9670da2, TFTP server pool exhaustion, authored 2 June and committed 27 October 2026, as GitHub prints itGitHub (lwIP)accessed 2026-11-16
- Reported byRelay of the report dated 16 November 2026: the 52% of the newly discovered, several projects have issued fixes. SecondaryHelp Net Securityaccessed 2026-11-16


