diff --git a/SECURITY.md b/SECURITY.md
index 33bbc2ff30..61e39361af 100644
--- a/SECURITY.md
+++ b/SECURITY.md
@@ -18,3 +18,120 @@ goes public.
 Only the maintained branches, as listed in [`BRANCHES.md`](BRANCHES.md),
 get security fixes.
 Users are urged to always use the latest version of a maintained branch.
+
+## Threat model
+
+We classify attacks based on the capabilities of the attacker.
+
+### Remote attacks
+
+In this section, we consider an attacker who can observe and modify data sent
+over the network. This includes observing the content and timing of individual
+packets, as well as suppressing or delaying legitimate messages, and injecting
+messages.
+
+Mbed TLS aims to fully protect against remote attacks and to enable the user
+application in providing full protection against remote attacks. Said
+protection is limited to providing security guarantees offered by the protocol
+being implemented. (For example Mbed TLS alone won't guarantee that the
+messages will arrive without delay, as the TLS protocol doesn't guarantee that
+either.)
+
+**Warning!** Block ciphers do not yet achieve full protection against attackers
+who can measure the timing of packets with sufficient precision. For details
+and workarounds see the [Block Ciphers](#block-ciphers) section.
+
+### Local attacks
+
+In this section, we consider an attacker who can run software on the same
+machine. The attacker has insufficient privileges to directly access Mbed TLS
+assets such as memory and files.
+
+#### Timing attacks
+
+The attacker is able to observe the timing of instructions executed by Mbed TLS
+by leveraging shared hardware that both Mbed TLS and the attacker have access
+to. Typical attack vectors include cache timings, memory bus contention and
+branch prediction.
+
+Mbed TLS provides limited protection against timing attacks. The cost of
+protecting against timing attacks widely varies depending on the granularity of
+the measurements and the noise present. Therefore the protection in Mbed TLS is
+limited. We are only aiming to provide protection against **publicly
+documented attack techniques**.
+
+As attacks keep improving, so does Mbed TLS's protection. Mbed TLS is moving
+towards a model of fully timing-invariant code, but has not reached this point
+yet.
+
+**Remark:** Timing information can be observed over the network or through
+physical side channels as well. Remote and physical timing attacks are covered
+in the [Remote attacks](remote-attacks) and [Physical
+attacks](physical-attacks) sections respectively.
+
+**Warning!** Block ciphers do not yet achieve full protection. For
+details and workarounds see the [Block Ciphers](#block-ciphers) section.
+
+#### Local non-timing side channels
+
+The attacker code running on the platform has access to some sensor capable of
+picking up information on the physical state of the hardware while Mbed TLS is
+running. This could for example be an analogue-to-digital converter on the
+platform that is located unfortunately enough to pick up the CPU noise.
+
+Mbed TLS doesn't make any security guarantees against local non-timing-based
+side channel attacks. If local non-timing attacks are present in a use case or
+a user application's threat model, they need to be mitigated by the platform.
+
+#### Local fault injection attacks
+
+Software running on the same hardware can affect the physical state of the
+device and introduce faults.
+
+Mbed TLS doesn't make any security guarantees against local fault injection
+attacks. If local fault injection attacks are present in a use case or a user
+application's threat model, they need to be mitigated by the platform.
+
+### Physical attacks
+
+In this section, we consider an attacker who has access to physical information
+about the hardware Mbed TLS is running on and/or can alter the physical state
+of the hardware (e.g. power analysis, radio emissions or fault injection).
+
+Mbed TLS doesn't make any security guarantees against physical attacks. If
+physical attacks are present in a use case or a user application's threat
+model, they need to be mitigated by physical countermeasures.
+
+### Caveats
+
+#### Out-of-scope countermeasures
+
+Mbed TLS has evolved organically and a well defined threat model hasn't always
+been present. Therefore, Mbed TLS might have countermeasures against attacks
+outside the above defined threat model.
+
+The presence of such countermeasures don't mean that Mbed TLS provides
+protection against a class of attacks outside of the above described threat
+model. Neither does it mean that the failure of such a countermeasure is
+considered a vulnerability.
+
+#### Block ciphers
+
+Currently there are four block ciphers in Mbed TLS: AES, CAMELLIA, ARIA and
+DES. The pure software implementation in Mbed TLS implementation uses lookup
+tables, which are vulnerable to timing attacks.
+
+These timing attacks can be physical, local or depending on network latency
+even a remote. The attacks can result in key recovery.
+
+**Workarounds:**
+
+- Turn on hardware acceleration for AES. This is supported only on selected
+  architectures and currently only available for AES. See configuration options
+  `MBEDTLS_AESCE_C`, `MBEDTLS_AESNI_C` and `MBEDTLS_PADLOCK_C` for details.
+- Add a secure alternative implementation (typically hardware acceleration) for
+  the vulnerable cipher. See the [Alternative Implementations
+Guide](docs/architecture/alternative-implementations.md) for more information.
+- Use cryptographic mechanisms that are not based on block ciphers. In
+  particular, for authenticated encryption, use ChaCha20/Poly1305 instead of
+  block cipher modes. For random generation, use HMAC\_DRBG instead of CTR\_DRBG.